Following exercise 2, it seems that the relationship between distance to kelp and density of zooplankton cannot appropriately be explained via a linear function. Therefore, I wanted to explore whether certain stations are correlated with one another to continue to try and figure out a way to create a spatial layer of zooplankton density across my entire study area that is ecologically informed. Therefore, my questions for exercise 3 were:
“Are there correlations between zooplankton density values at different stations?”
“If so, what is the nature of these correlations – positive/negative?”
“How does distance factor into these relationships?”
Approaches/tools used, methods & results
I first visualized the zooplankton density at all of my stations for 2018 across time (Fig 1). Since this plot looks a little convoluted, I also decided to split it out by the two sites (Mill Rocks and Tichenor Cove; Fig 2) as well as by station (Fig 3). Just by visual inspection, it does look like there may be peaks and troughs in zooplankton density occurring at the same time across several stations.
To investigate this, I then ran a bivariate correlation matrix on my data to compare prey density between all pairs of stations on the same days (Fig 4), which resulted in only 3 pairs being significantly correlated (Fig 5).
While this was interesting to see, these plots did not help elucidate where these correlations are occurring in space. Therefore, to add the spatial element into this, I plotted the correlation coefficients of each pair by the distance between that pair of stations (Fig 6). Looking at the plot, there is no obvious, discernable trend. Interestingly, there appear to be some stations that are strongly positively correlated but far apart, while there are also stations that are positively correlated that are close to each other. The same trend applies for negative correlations, though there is not as strong of a negative correlation as there is a positive one. However, a majority of the stations also fall in the middle, indicating that for a lot of stations, zooplankton density at one station does not relate to density at another station.
Although this plot did bring the distance between stations into context, it is hard to visualize that this actually looks like. Therefore, I visualized this plot in two ways. One was by drawing color-coded lines of the strongest positive and negative correlations on my site map (Fig 7), while the second was the same plot as Fig 6, however the points were color-coded by the habitat type (Fig 8) of the two stations (since habitat type is likely also a determinant of where we find zooplankton).
It has become very clear to me that trying to figure out a way to interpolate prey density across my sites is not as straightforward as I thought it was going to be, which can feel very frustrating at times. However, I simply need to remind myself that the marine environment is incredibly complex and dynamic and that it would be near miraculous if the relationships between habitat, environment, prey, and predator could be easily defined by one variable. For now, exercise 3 has continued to uncover the patterns that do and don’t exist in my data and has led to more analyses to try to keep disentangling the patterns. My next step for the final project will be to investigate temporal cross-correlation to see whether the density at station x1 at time t is related to the density at station x2at time t-1. I am a little doubtful about whether or not this will work because my sampling frequency varies between stations (sometimes we can’t do certain stations due to unworkable conditions) and there are sometimes 3 day gaps in sampling. However, I shall persevere and see how it goes!
Is Benthic Polychaetes Amphipods (BPA) ratio positively correlated to total organic carbon TOC off the Oregon and Washington coasts?
Is BPA ratio positively correlated to depth off the coast of Washington State?
How strong of predictors of BPA are TOC and Depth off the Washington coast?
Does the relationship between distance to nearest harbor/river and BPA differ at different depth ranges off the Oregon coast?
How effectively can you determine TOC based on depth and distance to nearest harbor or river in Oregon?
Tool or Approach Used:
Benthic Polychaetes Amphipods (BPA) Index (See Above)
Generate Near Table (Analysis) Tool in ArcMap
Join (Data Management) Tool in ArcMap
Multiple Regression Analysis w/Interaction Term
First, I queried by data in Access to determine the total number of polychaetae and amphipods in each box core sample. Then, I joined that information to and existing Excel file and generated a new column to determine BPA.
I typed the BPA equation listed above into Excel and used it to populate a new column to generate BPA index values for each benthic box core sample.
I created a shapefile that contained the GPS locations of all the major rivers outlets and harbors in Oregon.
I created a shapefile of all the benthic box core samples off the coast of Washington and then a shapefile of all the samples off the coast of Oregon.
I used the Generate Near Table (Analysis) Tool in ArcMap with Oregon benthic box core samples as the “Input features” and major river mouths & harbors in Oregon as the “Near features.” This generated a distance value from each box core sample to the nearest river or harbor.
I then used the Join tool too join the distance to nearest river or harbor values to the benthic box core data. After I completed the join, I exported the Attribute table to a text file and opened it in Excel using the “Add Data” function.
In Excel, I used the correlation and Data Analysis regression tools to answer the questions posed above.
Is Benthic Polychaetes Amphipods (BPA) ratio positively correlated to total organic carbon TOC off the Oregon and Washington coasts? – Washington: Pearson’s I Correlation Coefficient: 0.178069128 – WA Regression equation: y = 0.1332x + 0.803, y=BPA, x=TOC – WA Polynomial Relationship between TOC & BPA: y = -0.6365×3 + 0.5372×2 + 0.4564x + 0.7076
– Oregon: Pearson’s I Correlation Coefficient: 0.186643563 – OR Regression Equation: y = 0.0906x + 0.8261, y=BPA, x=TOC
2. Is BPA ratio positively correlated to depth off the coast of Washington? Oregon? – WA Pearson’s I Correlation Coefficient: 0.650932232 – WA Regression equation: y = 0.0079x + 0.3124, R2=0.4237, y=BPA, x=Depth
– OR Pearson’s: 0.179339185 – OR Regression: y = 0.0005x + 0.8135, R2=0.0322, y=BPA, x=Depth
-Because I saw a positive relationship between BPA and TOC from 0-200 m and a negative relationship at 200-600 m, I ran a Pearson’s I correlation coefficient on those two subsets of the depth data and found: – OR 0-200 m correlation: 0.504465898 – OR 200 -600 m correlation: -0.354042185
3. How strong of predictors of BPA are TOC and Depth off the Washington coast? – The results of a multiple regression analysis revealed an R2 value of 0.435745363315495 with y=BPA, x1=TOC, and x2=Depth.
4. Does the relationship between distance to nearest harbor/river and BPA differ at different depth ranges off the Oregon coast?
– Multiple Regression (at all depths): y = 0.760833428 – 0.000278547x1 + 0.741952481x2 -0.000206669, y= BPA, x1=Depth, x2= distance from river or harbor
5. How effectively can you determine TOC based on depth and distance from river or harbor in Oregon? – Multiple Regression: y = -0.141824199 + 0.003378366x1 + 1.147497178x2 + 0.003876666, y = TOC, x1=Depth, x2= Distance from river or harbor, R2 = 0.670711926
I found the Multiple Regression Analysis, Pearson’s I correlation coefficients, and ArcMap tools to be effective for my analysis. It was somewhat cumbersome to run these analyses in Excel, however. I would like to work with multiple regressions in R for ease of use, to determine interaction coefficients, and to graph results.
For the third exercise, I am exploring the MODIS fire data in more detail to access spatial patterns and seasonal impacts. I am still relying on mapping exercises and hotspot analysis to examine variations following the wheat harvest season in April and the rice harvest season in October. Some graphs are shown below:
In Exercise 2, I created a buffer zone and compared the relationship between road density and landslides. Exercise 3 I want to further analyze the impact of landslides on roads and traffic. For large landslides, how does it affect travel distance and time? If the landslides at the same location have different volumes, do people choose different ways to detour?
2. Tool or approach that be used
Buffer: First of all, attribute a table to find out what is volume and area of landslide, based on the area of landslide, determine if the construction site is needs, if yes, create buffer with appropriate distance.
Network analysis: Network analysis is to analyze what is trip time and distance with buffer. First click Network Analysis, then choose route as target. The route is presented, then click edit and create features, add stops and buffer, stops should be same when the route with or without buffer is analyzed.
Show directions: This tool is to estimate trip time and dictance. Also the difference between highways and urban roads will also be shown.
By comparing Figures 4 and 5, it can be seen that the 500m buffer zone will cause a 5-minute detour. Due to the difference in speed limit and traffic flow on highways and urban roads, the time may be longer.
Exercise 3: Visualizing the seedling structure and assessing seedling growth using point cloud data
The question that is addressing in this study
Characterization of the forest canopy and projection of forest inventory from Light detection and ranging (lidar) data has been widely used over the last two decades. Active remote sensing approaches (i.e., lidar) can use to construct three dimensional (3D) structures of tree canopies and underlying terrains (Pearse et al., 2019). However, due to some drawbacks associated with lidar data such as the low density, expensiveness of acquisition of data, and difficulties for utilizing for small areas, the practical use of lidar data becomes limited (Tao et al., 2011). However, the emergence of unmanned areal vehicles (UAVs) photogrammetric point cloud technology could able to address most of the drawbacks associated with lidar data. Especially, lightweight UAVs provide a better platform to acquire digital photos with lower operational costs (Ota et al., 2017). Therefore dense 3D point clouds generated from Structure-from-Motion (SFM) photogrammetry can use as ample source to substitute the lidar point cloud data and utilize to construct 3D structures to estimate forest biophysical properties.
Consequently, voxel-based matrics approaches are commonly used in forestry applications, especially for characterizing the forest canopy and other forest attributes in 3D space. With these approaches, point clouds are split “along both the vertical and horizontal axes to form volumetric pixels (volume pixel or voxel)” (Pearse et al., 2019). The main advantage of the voxel-based prediction is that we can extract the information within different layers of the forest canopy. However, most of the voxel-based matrices have been utilized with lidar point cloud data and terrestrial laser scanning data (Pearse et al., 2019). Since the acquisition of both lidar point cloud data and terrestrial laser scanning data expansive, UAVs based point cloud data use as an alternative solution. Therefore, in this exercise, I am interested in characterizing the forest attributes (in 3D space) using a voxel-based matrics approach by utilizing UAVs photogrammetric point cloud data. Specifically, identify “how the tree crown completeness varies as a function of the number of points that arranged in the 3D point cloud/ voxel resolution”.
2. Name of the tool or approach used, and steps followed.
Step 1: Pre-processing point cloud data
Selecting Area of Interest
I used R software packages to develop point cloud data from UAV images.
Packages used: lidR, sp, raster, rgdal, and EBImage
As the initial step, point cloud data were extracted using AgiSoft software (Fig.1(a)). Then the lasground function in thelidRpackage was used to classify the points as ground or not ground by using the “cloth simulation filter” (Fig.1 (b)). In the next step, I used the lasnormalize function in the lidR package to remove the topography from the point cloud created (Fig.1 (c)). Next, the normalized point clouds were cropped with the area of interest (AOI) shapefile (Fig. 2 (d)) using lasclip function.
Step 2: Voxelizing the point cloud data
For better visualization and to increase the efficiency of computer performance, I selected a subset of AOI to illustrate the voxel models(Fig. 1(d)). First, the “lasfilterduplicates” function (in lidR package) was used to remove the duplicate point from point cloud data. After removing the duplicate points, voxelization of point cloud data performed using the “lasvoxelize” function in the lidR package. This function allows two different options (cubic and non-cubic) to reduce the number of points by voxelizing the point cloud. Figure 2 represents the cubic voxel approach, and relative x, y, and z dimensions indicate the resolution of the voxel. Figure 3 represents the non-cubic approach, and relative dimensions for the voxel represent in each diagram. Especially in this approach, z-axis dimensions are different from x and y dimensions.
Step 3: Visualization approaches
The viewshed3d package uses 3D point cloud data for visualizing the different environmental settings, especially for construction 3D environments for ecological research. Due to the potential usefulness of this package, I used it for visualization of the study area and canopy structure of the seedlings in 3D space. At the initial stage, duplicated points were removed from the point cloud. Then the noise associated with point cloud data removed, and the reconstruct_ground() function was used to reconstruct the ground layer of the study area. Finally, an appropriate voxel model (described in section 3) used to visualize the seedlings in 3D space.
Step 4: Geographically weighted regression(continued from exercise 2)
Image classification tools available in ArcMap 10.7:
Constructing a digital elevation model
Spatial Analyst Tool: Tools available in Zonal
Geographically weighted regression
Assessing the effect of geography
To identify the influence of geography for seedling growth, I constructed a digital elevation model for the study area. Relative elevation values for each seedling location were extracted by using the Zonal tool available in ArcMap 10.7. As the next step, a geographically weighted regression performed by considering the tree height as the dependent variable and relative elevation of the seedling locations as the independent variable. The observed coefficients for the geographically weighted regression illustrated in figure 5.
Voxelization of point cloud data
Results obtain by voxelization of point cloud data indicate that reducing of voxel volume can provide better outputs in terms of visualization of seedlings in 3D space. For example, 1 x1 x 1 m3 cubic voxel cloud not able to visualize the structure of the seedlings in sub-AIO, while a volume of 0.1 x0.1x 0.1 m3 voxel could able to demonstrate the seedling architecture in an appropriate manner. After certain trials, a volume of 0.01 x 0.01 x0.01 m3 showed relatively the best voxel model for visualizing the seedling architecture. Even the voxel volume reduces (beyond 0.01 x 0.01 x0.01 m3), the visualization of the seeling article does not improve drastically. Further reducing the voxel volume requires more time and computer power to process the seedling article. Therefore, I selected 0.01 x 0.01 x0.01 m3 voxel volume (model) as the best cubic voxel model to identify the seedling architecture for this study.
Additionally, I performed voxelization of point cloud data with a non-cubic approach, and the observed results are similar to the cubic voxel models. The only difference for the non-cubic approach is the z-axis change (change of vertical resolution). We can change the resolution of the voxel image by changing the z-axis dimensions based on our interest. For example, we can characterize/observe the canopy structure by changing the resolution of the z-axis, as shown in figure 6. Non-cubic approach appropriate if we are interested in studying the variation of canopy structure relative vertically.
Based on the results obtained by visualization of the study area (step2), provide a better platform to understand the actual canopy structures of each seedling. Mainly, this helps to compare the results we obtained in the 2D environment (in exercise 2) and assess the results with the help of 3D representations. By looking at the canopy structures of the seedlings, we can confirm that seedlings growing in the northwest side of the AOI has relatively larger and healthy seedlings while the southeast side of the AOI has smaller and unhealthy seedlings. This observation agrees with the results obtained by geographically weighted regression (exercise 2 and 3). Both cubic and non-cubic approaches are applicable and can be utilized based on our desired output. The identified optimal voxel resolution (0.01 x 0.01 x0.01 m3) can be used as a reference for further voxel-based analysis, and it will help to save data processing time and with less computer power.
Results obtained by geographically weighted regression (height as the dependent variable and relative elevation of seedling locations as the independent variable) showed a decreasing trend of coefficient values with decreasing the elevation profile (Fig.4 (a)). Furthermore, obtained trends indicate that seedlings located in the northwest part of the study area have higher coeffects values, while seedlings located in the southeast part have relatively lower coefficient values (Fig.4(a)). Generally, the elevation of seedling locations may indicate essential parameters about the geography and water table of the study area. We can assume that lower elevation areas are more prone to accumulate water, and the groundwater table may close to the surface compare to the higher elevations. As described in the previous exercise, this might be the indirect cause for the presence of bare ground in lower elevated areas (i.e., stagnant water may damage the root systems of grass and may die due to rotten roots and will expose the bare ground). Similarly, the presence of excess water may hinder the growth of seedlings as well. As a side effect of the presence of groundwater with woody debris may tend to increase excess moisture conditions in subsurface soil layer and will act as a negative factor for seedling growth, especially for their root system.
Overall, the visualization of seedlings using the voxel approached showed promising results, especially identifying the optimal voxel size/volume that can be utilized for developing seedling architecture in future studies. However, to obtain better visualization outputs, we need to create high-density point clouds (described in step 1), and that requires more time and computer power.
If we are interested in a relative lager area, visualization of objects may need to be enhanced. Therefore, after selecting the appropriate voxel model (i.e., 0.001 x0.001 x 0.001 m3), plotting tools available in lidR package can be used to create better visualization structures, as illustrated in Figure A1.
Additionally, the detection of branches in early-stage seedlings may be difficult due to its unique shape and small branches. However, the developed approach can be useful for identifying the number of branches of a mature trees.
Ota, T., Ogawa, M., Mizoue, N., Fukumoto, K., Yoshida, S., 2017. Forest structure estimation from a UAV-Based photogrammetric point cloud in managed temperate coniferous forests. Forests 8, 343.
Pearse, G.D., Watt, M.S., Dash, J.P., Stone, C., Caccamo, G., 2019. Comparison of models describing forest inventory attributes using standard and voxel-based lidar predictors across a range of pulse densities. Int. J. Appl. Earth Obs. Geoinformation 78, 341–351. https://doi.org/10.1016/j.jag.2018.10.008.
Tao, W., Lei, Y., Mooney, P., 2011. Dense point cloud extraction from UAV captured images in forest area, in: Proceedings 2011 IEEE International Conference on Spatial Data Mining and Geographical Knowledge Services. Presented at the Proceedings 2011 IEEE International Conference on Spatial Data Mining and Geographical Knowledge Services, pp. 389–392. https://doi.org/10.1109/ICSDM.2011.5969071.
Mineral dust is the most important external source of phosphorus (P), a key nutrient controlling phytoplankton productivity and carbon uptake, to the offshore ocean (Stockdale et al., 2016). Paytan and MacLaughin (2007) emphasized that atmospheric P can be important as the major external supply to the offshore ocean, particularly in oligotrophic areas of the open ocean and areas that are P-limited, such as Bermuda Ocean. The most important source of atmospheric P is desert dust, which has been estimated to supply 83% (1.15 Tg⋅a−1) of the total global sources of atmospheric P (Mahowald et al., 2008). Of that dust, it is estimated that 10% is leachable P (Stockdale et al., 2016). In addition, Saharan dust supplies a significant fraction of the P budget of the highly weathered soils of America’s tropical forests and of the oligotrophic water of the Atlantic Ocean, increasing the fertility of these ecosystems (Gross et al., 2015).
Hence, this background is my starting point to try analyzing the correlation between the Particulate Organic Phosphorus (POP) and Primary Production (PP) in Bermuda Ocean. In exercise 2, I tried to answer these questions, however, the result was not likely that I expected. In this exercise 3, I explored a lot about my raw data and I realized that the way I divide the data will affect the result a lot. In this exercise I dug up a lot about the time series, regression, and auto-correlation function in R.
However, due to my misinterpret data in exercise 2, I ended up adding one more variable in this exercise 3 to broaden my analysis and convince my result. The previous study revealed that most phosphorus (P) and iron (Fe) are present as minerals that are not immediately soluble in water, hence not bioavailable (Lidewijde et al (2000), Shi et al (2012)). Lidewijde et al (2000) also stated that phosphorus (P) and iron (Fe), if deposited to the surface ocean, may pass through the photic zone with no effect on primary productivity, owing to their high settling velocity and low solubility. The photic zone has relatively low levels of nutrient concentrations, as a result, phytoplankton does not receive enough nutrients. Moreover, there are several factors that contribute to the primary production, such as physical factors (temperature, hydrostatic pressure, turbulent mixing), chemical factors (oxygen and trace elements), and biological factors. Hence, I added the temperature variable in this exercise 3.
In this exercise 3, I would like to find out what factor that contributes to the PP in Bermuda Ocean and how is the time cycle of three variables (PP, POP, and Temperature)?
Dr. Julia helped me in pre-processing data by using excel, and for further analysis, I used three tools in R:
Time series function: This function will help us to plot the time series trend for every data. Ex code: plot.ts(POPR['PP'], main="PP depth 0-7 meter", ylab="PP")
Linear regression model: This function will help us to identify the correlation between two designated variables, in this exercise, the dependent variable is Primary Production and the independent variables are POP and Temperature. Ex code: POPR.lm<-lm(POP~PP,data=POPR) and summary(POPR.lm). To plot it into a scatter and linear line, I used ggplot function. Ex code: ggplot(POPR, aes(x=POP, y=PP))+ geom_point() + geom_smooth(method=lm,se=FALSE)
Auto-correlation function: This function is used to find patterns in the data. Specifically, the autocorrelation function tells you the correlation between points separated by various time lags. In this exercise, the lag ranges from +1 to -1, where +1 is perfectly related and -1 is inversely related. In afc chart, the dashed line represents the boundary of the significance of correlation. Ex code: acf(POPR['PP'],lag.max = 51,type = c("correlation", "covariance", "partial"),plot = TRUE,na.action = na.contiguous,demean = TRUE) In this code, we can change the lag.max to the number of the data if we want to access the lag coefficient individually, or uses NULL to default. For this exercise I used lag.max which according to the number of my data.
Steps of Analysis
Divided the data based on the depth categories:
Table 1. The Category of POP based on depth
2. Regression analysis for between PP vs POP and PP vs Temperature. Since the data of PP and temperature is only for 0-8 meter depth, hence the regression is only conducted for in this depth category. In order to perform the regression, I pulled out four outliers:
3. Accessing the temporal pattern of variable PP, temperature, and POP by using the auto-correlation function and also time series function. Due to some missingness data, especially for POP and temperature in 0-7 meter depth, hence the date 2014/03/06, 2015/02/04, 2015/12/13 is pulled out for POP and date 2014/03/06 and 2014/12/11 is pulled out for temperature.
From figure 1, overall the POP has a positive correlation to the PP, and the temperature has a negative correlation to POP (R square PP vs POP is 0.3064 and R square PP vs temperature is -0.183). From this analysis I can assume that POP is a factor that contributes to the Primary Production in Bermuda ocean. However, to see the detail of this analysis I tried to see the regression analysis per month in a 5 year interval period. I did not perform the regression in January due to a lack of data. From figure 2, we can see that from April to December and February, the pattern of correlation between PP vs POP is similar to the pattern of PP vs temperature, where the highest is in May and the lowest is in December and February. The significant difference happens in March, where the gap is +0.98 for PP vs POP and -0.99 for PP vs Temperature.
2. Auto-correlation and Time Series PP, POP and Temperature depth 0-7 meter
From figure 3, as we can see the temporal pattern of PP and POP is likely similar, where the highest peak is at the beginning of the year for every year, except for 2012 and 2016. This is due to in 2012 the data start in June and March in 2016. This pattern is inversely for temperature, wherein the beginning of the year the temperature is very low and high from June to July. To confirm this pattern, we can see the autocorrelation (ACF) chart, where the temperature has significantly related to the function of time, The vertical line crosses the horizontal dash line which means there is a repetition cycle in time for temperature.
As a temporal pattern, there is a repetition pattern for POP and PP, where the highest value happens in January-March. However the difference from temperature is the repetition in temperature has the exact same value, which does not happen for POP and PP. There is a big gap between the value of POP and PP in January and February from 2013 to January and February in other years. I think that is why there is only one significant correlation line in the beginning ACF chart for POP and PP.
3.Auto-correlation and Time Series of POP by depth categories
From figure 4 we can see that the temporal pattern from lag 0-5 is similar for all depth categories where for overall lag only POP 1 to 3 is similar (depth 0-22 meter). It indicates that the deposition of the POP can reach 22 meters depth at the same time.
If we see at POP 4 to POP 6, there is no repetitive pattern both in the ACF chart and the time series chart. However the repetitive pattern happens from POP 7 to POP 10, or from 98-164 meter depth.
By looking at the number of POP 11 (depth 197-200 meter), there is only a small number of POP can reach this depth.
Critique of the method – what was useful, what was not?
Overall this exercise 3 made me realized the way we process the data will affect the result. In exercise 2, I was likely to simplify the process, hence I got difficulties in order to interpret my data and the result is not likely what I expected.
In this exercise 3 I learned a lot about the Auto-correlation function and time series function in R. In my opinion, acf is very useful for stable data like temperature, where the repetitive pattern is along with the value. The significance of ACF is based on the value, hence in the ACF temperature chart (figure 3) we can see that most ACF has a significant correlation overtime period (both positive and negative), which does not happen in POP and PP data.
POP and PP have a pattern, however, the value varies overtime period, hence although we can see the pattern in time-series and ACF charts, the significance only appears in the beginning month in 2013. In my opinion, if we would like to access the significance cycle, the unstable data like POP and PP where the is a big gap value overtime period, ACF is not really suitable. However, if we just would like to see the pattern we can relly on time-series and ACF function as well.
Gross, A., Goren, T., Pio, C., Cardoso, J., Tirosh, O., Todd, M. C., Rosenfeld, D., Weiner, T., Custódio, D., & Angert, A. (2015). Variability in Sources and Concentrations of Saharan Dust Phosphorus over the Atlantic Ocean. Environmental Science & Technology Letters, 2(2), 31–37. https://doi.org/10.1021/ez500399z
Eijsink LM, Krom MD, Herut B (2000) Speciation and burial flux of phosphorus in the surface sediments of the eastern Mediterranean. Am J Sci 300(6):483–503. doi: 10.2475/ajs.300.6.483.
Mahowald, N., Jickells, T. D., Baker, A. R., Artaxo, P., Benitez-Nelson, C. R., Bergametti, G., Bond, T. C., Chen, Y., Cohen, D. D., Herut, B., Kubilay, N., Losno, R., Luo, C., Maenhaut, W., McGee, K. A., Okin, G. S., Siefert, R. L., & Tsukuda, S. (2008). Global distribution of atmospheric phosphorus sources, concentrations and deposition rates, and anthropogenic impacts. Global Biogeochemical Cycles, 22(4). https://doi.org/10.1029/2008GB003240
Stockdale, A., Krom, M. D., Mortimer, R. J. G., Benning, L. G., Carslaw, K. S., Herbert, R. J., Shi, Z., Myriokefalitakis, S., Kanakidou, M., & Nenes, A. (2016). Understanding the nature of atmospheric acid processing of mineral dusts in supplying bioavailable phosphorus to the oceans. Proceedings of the National Academy of Sciences, 113(51), 14639. https://doi.org/10.1073/pnas.1608136113
Zongbo Shi, Michael D. Krom, Timothy D. Jickells, Steeve Bonneville, Kenneth S. Carslaw, Nikos Mihalopoulos, Alex R. Baker, Liane G. Benning. Impacts on iron solubility in the mineral dust by processes in the source region and the atmosphere: A review. Aeolian Research, Volume 5, 2012, Pages 21-42. ISSN 1875-9637. https://doi.org/10.1016/j.aeolia.2012.03.001
For this exercise I was interested in continuing to test out different variable combinations with in the Geographically Weighted Regression and further exploring the meaning of the relationship between these variables. This framed exercise three with the following research questions:
What are the individual relationships between change in race, education, and income across both decades?
How does each change variable relate to itself across the two decades?
Do these results produce information that provide information on where there might be gentrification?
How do these results compare to where others have identified gentrification and how they have calculated it?
Geographically Weighted Regression
I continued using the GWR tool to make more maps of how the variables are correlated. I expected some similarity in the patterns between variables, and I hypothesized that they would be in the northern central part of the city. When I compare each decade’s variables to the same variable in the second decade, I expect some correlation in areas that have steady change, but lower correlation in areas that change rapidly in one decade and not the next.
In order to do this, I used the GWR tool in ArcGIS and used the number of neighbors as the neighborhood type. I ran the tool with change in race from 1990-2000 as the dependent variable and change in education from the same decade as the explanatory variable. I repeated this with race and income, and then all of it with the second decade. I then ran it with education as the dependent variable and income as the explanatory variable for each decade.
Comparison to Other Gentrification Maps
I have begun to look at other maps of gentrification in Portland, OR to help make sense of the information that I am getting out of the GWR tool, compare results, and compare methods. Bureau of Planning and Sustainability for the City of Portland has produced maps of gentrification, also by census tract throughout the city. They created a taxonomy of gentrification from susceptible to late gentrification (and even continued loss after gentrification). They also have maps of the variables that went into their maps, which consisted of percentages of populations that were: renters, of color, without a bachelor’s degree, and how many households were below 80% of the median family income. These were all combined into a vulnerability map. The changes in renters, race, education, and income were combined to create a map of demographic change. The change in median home value and appreciation from 1990-2010 maps were combined to create a housing market conditions map. All of these were combined to create the final gentrification map.
The results of the GWR tool produced maps comparing two variables for one decade. As mentioned above, I ran the tool for all variable combinations for both decades. These results show that each combination of variables changes together in a different part of the city. Change in race and education occurs in the northwest part of the city in the first decade and north central/west part in the second decade. Race and education change in the eastern part of the city in the first decade, and some in the east and some in the northwest in the second decade. Education and income change in the southern part of the city in the first decade and southeast in the second decade. It is interesting that all of the combinations are slightly different but have some patterns.
The regression coefficients from the model runs when comparing each variable across the two decades show interesting patterns too. The changes in race each decade was relatively similar across the city but had the highest coefficients in the southern part of the city. Changes in education both was the same in the eastern part of the city, and changes in income both decades were similar in the north/central part of the city.
When compared with the Bureau of Planning and Sustainability gentrification maps, there are some patterns of similarity. They find that the northwest part of the city there is gentrification currently happening, the central part of the city has a late stage of gentrification, the south part has some early stage gentrification, and the eastern side is susceptible.
The correlation between the changes in race and education seem to have similar patterns to current gentrification and areas that are susceptible. Change in race and income seem to follow a similar pattern to susceptible areas. Change in education and income seem to have the highest correlation in areas that are susceptible or are in early stages of gentrification.
I think the biggest difference in their calculations are using the opposite side of the data. For race they used percentage of populations that were communities of color, whereas I used percentage that is white. They used percentage without a bachelor’s degree, whereas I used percentage with a bachelor’s degree. They also used the percentage of the population that are renters. These differences tell a slightly different story since there are multiple patterns of change in the city of Portland.
I think one downfall of the GWR tool is that if you include multiple explanatory variables, you cannot get a coefficient value for how they all change together, you get separate coefficients for each explanatory variable. Granted, I’m not sure how this would be possible, but it would be useful to see if the three variables have any patterns together. This makes it difficult to compare three variables at the same time.
# Lewis et al. 2007, D’eon and Delparte 2005. data screening based on fix type and PDOP (DOP)
# calculate data retention for: (Blair 2019)
# 1. Converting all 2D locations to “missed fix”
# 2. Converting all 2D locations AND 3D with DOP>5 to “missed fix”
# 3. Converting all 2D locations AND 3D with DOP>7 to “missed fix”
# 4. Converting all 2D with DOP >5 AND 3D with DOP >10 to “missed fix”
To account for fixes that are likely inaccurate (not enough satellites), there are various ways to screen the data. I tried 4 different methods to see which method had the highest data retention. Fixes that are considered inaccurate past a certain threshold are flagged as a “missed fix” even if the collar obtained a location.
Comparison of data retention given 4 different screening methods, versus no screening method. The original proportion of successful fixes out of the total number of attempted fixes was 0.91, and the next highest was method 4 (Removing all 2D with DOP >5 AND 3D with DOP >10).
Option 4 (Removing all 2D with DOP >5 AND 3D with DOP >10) had the highest data retention after removing suspicious locations. Method 1 (removing all 2D locations) was a closet second.
The next step is to run the models for FSR as a function of environmental attributes again with the new screened data (derived from Method 4).
ANNULUSMOVING WINDOW ANALYSIS
First attempt at moving window analysis used very large window sizes.
The effect of sky availability at various spatial scales on FSR. All models were competitive.
Decided to try again with smaller window sizes.
05/22/2020 finished redoing window sizes, next need to extract values from test sites and re-run the regression analysis.
Context: I calculated the difference between each test site location (obtained via handheld GPS unit) and collar locations. Common data screening practices utilize cutoff values based off of fix type (2D vs 3D fixes) and Dilution of Precision (DOP), which are both recorded with the coordinates obtained by the GPS collar. Locations predicted to be more accurate (less location error) are normally 3D, and have low DOP values.
A common practice is to remove all 2D locations, or remove all locations with DOP>5.
The graphs below are meant to help me figure out a cutoff value for the stationary collar test data. The cutoff value used here will also be applied to the deer GPS data. The graphs are the result of a series of subsetting the data based off of DOP values and Fix Type to help gain a clearer picture of the data.
I was hoping to see a more clear pattern or “threshold” pattern that would show that above a certain DOP value, distances dramatically increased. However as the “data cloud” in the first scatterplot shows, there isn’t much of a pattern. However…
If we look at only the locations within 30 meters to match raster data resolution, we see that this data includes DOP as high as 18.5 and does not include any 2D data. The fact that these more accurate locations also include such high DOP values is unexpected. I’m not sure what the best cutoff value is based off of these plots.
For exercise 3, I am interested in the relationship between accuracy of wood detection using a random forest classifier and distance from training samples used to train the classifier. Specifically, do accuracy metrics (i.e. accuracy, kappa, Type I and Type II error) change with distance to training area? To train the classifier, I selected a 3.5-acre area near the center of my area of interest. I selected this area because I felt it had a variety of features—islands, wood jams, shadows, woody debris, rocks, etc.—representative of features found throughout the 150 acres of the site. I hypothesize that kappa will decrease in sample plots that are further from the training area.
I have not executed the approach I intend to use because I am still in the process of data wrangling and creating confusion matrices for sample plots (I am attempting to batch clip plots in ArcGIS Pro, and merge rasters into a 2-band raster stack with reference data in one band and predictions in the other band), however, I ultimately intend to use a similar approach as I implemented in exercise 2 using simple linear regression in RStudio and geographically weighted regression in ArcGIS Pro.
I will calculate distance from sample plot centers to training area using the near tool in ArcGIS Pro. This distance will be the explanatory variable in a linear model, and kappa will be the dependent variable. I would like to append this data in tabular format as illustrated in Table 1 below.
Distance to Train Area
I will calculate kappa and distance to training area (ideally for 72 sampled plots). Then, I will use these variables in a simple linear regression in RStudio and a Geographically Weighted Regression in ArcGIS Pro.
4. I anticipate the results of this analysis will indicate a decrease in kappa as plots are located further form the training area. I hope to see this indicated by a low p-value in SLR. The GWR may suggest there are areas of the site where the classifier performed poorly, and I believe this will be likely due to lighting conditions in the area (either shadows, or very high reflectance on woody debris due to sun angle). If there is time, I would like to assess the relationship between kappa and wood area because I currently cite Landis & Koch 1977 to assess kapa values. This paper refers to a somewhat arbitrary scale for assessing the accuracy of a classifier based on kappa. It would be interesting to see the effect kappa has on wood area estimates.
5. I will critique the method when I have results. My early impressions are this methodology is a little tedious due to number of plots. I am having a lab assistant manually delineate the wood polygons in plots that were also sampled on the ground. These plots will then be rasterized and clipped. I also need to clip the classified raster by these plots (there is a batch clip tool available in arc). From here, I will write a loop in R that accepts a path, locates tiffs in the folder, and calculates confusion matrix metrics for each plot.
Landis, J. R. and Koch, G. G. 1977. The Measurement of Observer Agreement for Categorical Data. Biometrics, 33(1): 159. doi: 10.2307/2529310
For this week’s exercise I was interested in refining my spatial model specifications based on my findings of previous weeks. I was interested in asking the question:
Can I determine a spatial linear model to describe the relationship between county out migration flows and flood claim payments and claim rates?
Significance in context: Modeling the relationship between county migration flows and flood claims allows for discussions of the implications of flood insurance on population mobility and how natural hazards interact with the US population landscape in general.
2. Name of the tool or approach that you used.
In order to test the fit of two model frameworks (the spatial lag and spatial error model) I used the Lagrange Multiplier Test which tests spatial dependence in model fit for these two models under typical and robust forms. I then fit model variables to the preferred model form using the “spatialreg” package in R and map the residuals to observe spatial autocorrelation.
3. Brief description of steps you followed to complete the analysis.
I started by subsetting my data into only counties in Gulf Coast states (Texas, Louisiana, Mississippi,Alabama, and Florida). I then created 2 different versions of this subset of states for a high flood instance year (2005) and a low flood instance year (2003).
Next, I created an Ordinary Least Squares model where log base 10 county out-flow is a function of log base 10 total amount paid and claim rate per county (log(out_flow) ~ log(total_amount_paid) + claim_rate). Following this, I ran a Lagrange Multiplier Test and used the outputs to determine the preferred alternative model based on spatial dependence. I applied the following framework created by Luc Anselin (2005) to determine the preferred alternative model.
I then proceeded in running the preferred model, which, in my case was the spatial error model. The model structure is composed of the same variables described in the OLS version. I then map the residuals by county of the error model and the residuals of the original OLS model for comparison.
4. Brief description of results you obtained.
The spatial error model results can be seen below with mapped residuals alongside for 2005 and 2003.
The OLS mapped residuals can be seen below for comparison for 2005 only.
These results indicate there is statistical relationship in insurance payment amounts that is positively associated with county out flows. This means that as payment amounts increase we expect there to be more people leaving. Interestingly enough there was a notable difference in 2005 and 2003 for the coefficient estimate associated claim rate and I would attribute this to the relatively lower claim rate in 2003 compared with 2005. Next, I will be interested to see if this negative (not statistically significant) relationship remains when I expand the model and test for all years. Notably, we see the model is much improved in spatial dependence by using the spatial error model which can be seen comparing the residuals mapped in the OLS map and the Error maps.
5. Critique of the method – what was useful, what was not?
This method was very useful in helping me find a model specification that more precisely fits the data without spatial dependence. I will have to determine if the same model structure is viable at a larger scale spatially and temporally next.
The only cons of this approach is the extensive research required before understanding the data and diagnostic test results well enough to proceed and communicate the interpretations.
I initiated exercise 3 expecting to glean useful information from GWR and global Moran’s I experiments, but preliminary exploration of these tools indicated that they wouldn’t be particularly useful to me given how my data is currently set up. As a result, I decided to go ahead and attempt my own version of a network analysis, to better characterize ice velocity around the borders of Greenland. In exercise 2, I simply pulled velocity data from right by glacier termini- which is not necessarily the be all and end all of glacier velocity. My time is better spent trying to develop a more robust velocity dataset, and returning to more conventional techniques for interpreting statistical relationships. Furthermore, I went to great lengths to improve data quality (fewer glaciers with gaps in their datasets…) in hopes that this manifest in clearer statistical relationships.
1)How does velocity along entire Greenland ICE FLOWS (or ice streams) correlate to terminus retreat rates for their associated outlet glaciers?
In simpler terms: Do glaciers that are fed by larger/ faster ice flows display a greater likelihood of rapid retreat?
2) Does the relative annual change in ice flow velocity (variability in %) correlate to relative annual change in glacier retreat (variability in %)?
In simpler terms: if ice velocity in a location jumps from one year to the next, do we see a greater likelihood of terminus retreat increasing as well?
Given time constraints for the class, and my particular questions, I had to use some intuition to rule out various paths forward. Exploratory statistics in my velocity and retreat datasets that there is really no relevant geospatial clustering. (sure, high velocity clusters around valleys, but this is where all the outlet glaciers occur anyways, so that is relatively meaningless/ obvious). It seems clear to me that I need to better represent velocity data, on a wider spatial scale, in order to accurately represent the relationship between ice velocity and terminus retreat. To do this, I attempted a rudimentary “network analysis” to incorporate velocity data along entire flow lines of ice. This seems far more valuable to me than fiddling with meaningless GWR and correlogram data on relatively messy datasets.
APPROACH: “Characterizing Ice Stream Velocity”:
ArcGIS Pro Tools used:
Drawing polylines, Generate Points along Lines, Extract Multivalues to Points, Summary Statistics, Join Field, Buffer, Intersect, Table to Excel
Excel Tools used:
Time series, R- squared simple regressions, normalization, first derivatives
1.DRAWING FLOW LINES: A feature class was created and populated with flow lines. Flow lines were drawn using the feature editor, and using visual queues from velocity datasets. Care was taken to ensued flow lines spanned the entire study period (now 2008-2016) so as not to have gaps in my dataset. If glaciers did not appear to be associated with ice flows, fell within a velocity satellite “dead zone” or did not have terminus data for most years, they were excluded from the dataset. The total 240 glacier dataset was shrunk down to 140 glaciers following these limitations.
2)GENERATE POINTS ALONG LINES tool was used to split each flow line into 20 parts (5% chunks). Each point could then be used to extract underlying velocity data from annual velocity rasters.
3)MULTIVALUES TO POINTS tool was used to extract velocity data from velocity rasters spanning the study period. This data is saved in the point attribute table.
4)SUMMARY STATISTICS tool was used to summarize these velocities. For each glacier, the MEAN and the SUM of velocity along flow line points was recorded.
5)JOIN FIELD tool was used to apply these new MEAN and SUM variables to the flow line attribute table.
6)BUFFER tool was used to expand previous glacier point data representing glacier ID, as well as previously calculated terminus retreat data over the study period.
7)INTERSECT tool was used to merge the flow line dataset (containing new velocity data) with the glacier point dataset (containing old retreat data). This was to consolidate data into one attribute table.
8)TABLE TO EXCEL tool was used to export the attribute table for use in other programs.
An exhaustive array of graphs and statistical relationships were generated in order to try to understand how ice velocity and terminus retreat interact, but I will share only the most insightful in this blog post. For the most part, the results are very similar to those from exercise 2, which isolated velocity data to point sources right near termini.
When averaged across the entire island, there appears to be a relatively strong correlation between the average ice flow velocity and the average terminus retreat, indicating that these two variables are indeed connected on an annual timescale. However, considering there are only 5 discontinuous years to plot, this conclusion would benefit from a greater pool of sample years. With an R-squared of 0.63, we can with large uncertainty say that on average, years that experience faster glacier velocity in ice streams also experience somewhat faster retreat at their associated glacier termini. This is a very similar conclusion (with a very similar R-squared) to what was reached in exercise 2 using point-sourced velocity data. It is worth noting that from 2008 to the 2016-2017 season, there does not appear to be an increasing trend in terminus retreat or in velocity along flow lines, which does not follow local temperature trends. This indicates a potential lag in response time larger than that of the study period (not at all surprising).
In addition to plotting the absolute magnitudes of these two variables, I wanted to plot the temporal change. This was a bit unorthodoxed given the discontinuous nature of the sample years, but it still provides valuable information. Again averaged over the entire island we see another trend: Glaciers that experience an increase in velocity from t1 to t2 are more likely to increase their rate of terminus retreat from t1 to t2. These variables can be conceptualized as the first derivatives of the records provided above- focusing on the rate of change rather than the magnitude of change itself.
While annual data averaged over the entire island showed clear trends, much like it did last time, individual glaciers for every given sample year tell a different story. Not displayed here: there was no trend between absolute velocity along flow lines (MEAN OR SUM) and absolute distance retreated in any given year. As you can see above, there is also virtually no trend between their first derivatives either. Glaciers with bigger jumps in velocity from t1 to t2 do not necessarily experience greater jumps in rate of retreat from t1 to t2. This goes against my initial hypothesis from the start of the project (that faster moving glaciers also retreat faster.
Normalizing these changes over time also does not induce a visible trend. The figure above is merely an example figure, but all four time jumps show no clear trend even on a relative scale. However these regression plots could theoretically tell us something about variability in any given year if the sample periods were better represented (consistent time jumps). For example, glacier velocity seems to have been much more variable in the 2008-2014 period, compared to 2014-2016. However, this is not a fair comparison, because the first 2 show a change over several years, whereas the second 2 are actual year to year changes. (its also not fair to divide the first 2 over their study period, because that assumes the changes were incremental and constant, which they most certainly were not.)
CRITIQUE & FINAL STEPS
The problem with my approach is that I do not have the necessary data to draw a reasonable conclusion. There are just far too many temporal gaps in the dataset(which span too short of a time period to observe trends in glaciers- which have notoriously slow response times). I did my best to eliminate glaciers that did not have all the necessary data (whittling down from 240 glaciers to 113 based on available data), but I cannot generate data for years that were simply not provided. As a result, my conclusions can really only be as comprehensive as the data allows.
As Julia brought up in my previous draft, my hypothesis also applies to a specific subset of glaciers- those with wet basal conditions. It is prudent to presume that moving forward, more and more outlet glaciers will develop these conditions, but once again, glaciers are slow to respond to climate responses. It is also quite hard to find a dataset that determines which of these glaciers are wet-based vs. dry based: sure we know for a fact that certain glaciers may or may not be wet-based, but glaciers where these conditions are well constrained are not just scattered across the continent. One idea for a future project would be to limit this analysis only to glaciers we know have wet basal conditions. As of right now, it seems those are the only glaciers velocity might have predictive power over to begin with.
Finally, the hypothesis can simply be wrong. I was operating based on an assumption that vulnerable glaciers tend to heat up and develop wet beds- which cause them to flow faster. These same glaciers would hypothetically also be more vulnerable to retreat at the terminus because they have an assumed negative mass balance. An alternative mechanism- as Julia pointed out, could be that glaciers that FLOW faster and are indeed melting might not necessarily exhibit terminus retreat, because that lost “real estate” at the terminus would be quickly replenished by new ice from further inland. That too is a very valid hypothetical mechanism.
Moving forward, I think the only thing I could milk from this dataset without starting an entirely new project is to try to identify regional trends. I cannot distinguish glaciers based on their basal conditions with the available data, but I CAN distinguish them based on their location. Perhaps the different regions of Greenland glaciers are experiencing different trends. The final step I will complete will be to divide the data I produced above into 6 separate regional groups, indicated below (over GRACE- mass anomaly data) and recreate the graphs above to distinguish any trends in different regions, if there are any present. It could be reasonable to assume that some of these regions might have higher proportions of wet beds than others.
Up until this point, I have worked through 2-Dimensional representations of the artifact assemblage from Cooper’s Ferry, Idaho. My first blog post contained visualizations of the artifacts within 3-Dimensions and 2-Dimensions by using kernel density. These graphics were used to identify intervals at this archaeological site that appear to be separated by an extended period of sediment deposition and/or time. Under the hypothesis that human behavior at the site differs between these “gaps” in time, I chose five unique intervals to examine. Within a 1-meter depth, I chose the intervals 0-25cm, 25-50cm, 50-60cm, 60-74cm, and 74-100cm. Geoff Bailey (2007), when discussing time perspectivism, states that archaeological assemblages may be best studied through multiple scopes in time and space. The idea represented here alludes to a much more formalized theory in fractals (Mandlebrot 1967; Plotnick 1986; Sadler 1999). In a general sense, a fractal is a pattern that is repeated at different scales. Mandlebrot (1967) discusses how fractals can be used to define the coastline of Britain at different spatial scales. The larger scale uses fewer fractals, or lines, to define the coastline of Britain, whereas a smaller scale requires more fractals because there is more identifiable detail in the coastline when one “zooms in” on a map. My question for this blog post mirrors this discussion on fractals; the main difference in my study is that, instead of looking at coast lines, I wish to understand how the temporal pattern can be observed within the archaeological assemblage between and within each interval listed above. To explore this inquiry, I used a model developed by Maarten Blaauw and J. Christen that has been named after Francis Bacon (AD 1561-1626).
Question:How is sediment accumulation rate associated with artifact deposition through different temporal scales at the archaeological site in Cooper’s Ferry, Idaho and how can these rates of deposition imply natural and cultural transformations on archaeological deposits?
For the purpose of thinking about how the artifacts in this assemblage are separated by time, I explored the concept of age-depth models that I was directed to with the assistance of Erin Peck. Age-depth models, with the proper input, facilitate the understanding of depositional rates in sedimentation throughout a core of sediment. The primary input for the depth-age model I used is C14 dated materials that was recovered from the archaeological context in question. I utilized the ‘rbacon’ package built for the R environment. Following the guidance and recommendations given by Blaauw et al (2018), I implemented the carbon dates I had access to (fig. 1) and produced an age-depth model correlating to those known dates (fig. 2).
age cal BP
Figure 1. This chart shows the 8 C14 dates that have been collected from the context of the current archaeological assemblage. Dates and associated errors are given in calibrate years (cal) Before Present (BP) at the related depth the material was recovered (cm).
How to Cook the Bacon
There is a large number of components associated with creating an age-depth model within the ‘rbacon’ package. I have explored multiple different approaches with certain settings set to ‘on’ and others set to ‘off’. Now, one can simply add the dates recovered from the context in question and run the program, but there is a lot of information to think about and consider implementing to produce stronger and more accurate age-depths. The main algorithm behind this model production (fig. 2) consists of Bayesian statistics which takes prior knowledge and information in account for the improvement of inferential statistical analyses (Blaauw et al 2018). One of the most important inclusions to consider is known or hypothesized hiatuses. [I am sure you were waiting until fractals came back into conversation] Hiatuses have been shown by Sadler (1999) and Plotnick (1986) to create a process of temporary deposition followed by an erosional event. A hiatus can be thought of as a period of no sedimentation, but it can also lead to a net loss in sediment via erosional events. When these patterns appear in the sediment record at small time frames it is understood through a net-gain/net-loss system. When the time frame is broadened, the pattern identified in the record appears to represent a net-gain and net-loss sedimentation rate of zero. When this model is plotted on a graph, the figure resembles stairs when the x-axis represents time and the y-axis represents depositional rates. Thus, the horizontal segments of the stairs resemble a period of no deposition within a broad scale of time, henceforth a hiatus is brought into the picture [and fractals are born]. In the archaeological sciences, Schiffer (1987) hypothesizes that these types of events can displace and reposition cultural materials through both natural and cultural transformation of the area. Therefore, it is important to implement as many known or hypothesized dates for a hiatus or other surface altering events (i.e. volcanic eruptions, landslides, etc). Unfortunately, there is no known or hypothesized hiatus within the context of this archaeological assemblage because it requires a more in-depth look at the geological processes during this time in the region; something that might be great for future work and interpretation.
Thinking about hiatuses in terms of sedimentation and cultural activity are a separate, yet intriguing, concept to consider in this instance. A hiatus in sedimentation can often represent relative stability in the land form and allow for a pedostratigraphic horizon to begin forming. Pedostratigraphic layers, from a geological view, are horizons within sediment that allow soils and organic materials to form. [People really like soil and do not always enjoy rocky or sandy surfaces] The presence of pedostratigraphic soils often represent ancient surfaces that have been buried by the sediment of the past 6 million days. These ancient surfaces have a high probability of containing cultural materials because soils seem to attract human activities. To summarize, when we have a hiatus in sedimentation, we have stability in the land form, so we get soil formation, and henceforth a lot of human activity. [In a simple format, sediment hiatus = lots of people] On the other hand, a hiatus in human activity can sometimes, not always, represent instability in the land form and thus, rapid sedimentation. [From this, high sedimentation rate = minimal rates of human activity] Now, can dense human activity/occupation indicate a hiatus in sedimentation, and can rapid sedimentation predict a hiatus in human activity? The result from the age-depth model production will be discussed in the next section and I will briefly explain how the scope and time frame of investigation [spoiler alert: the model does not seem viable to answer my question] needs to be much tighter than the current results to determine how sedimentation at this site affected human activity.
What does the Bacon look like When it is Done?
As you can see in figure 2, there is a pretty straightforward linear relationship between the dates I provided and the predicted depth per time interval. As previously mentioned, the inclusion of many more dates would further define the sedimentation rate through time (Blaauw et al 2018). Unfortunately, the process of collecting further dates may be beyond the scope [pun intended] of my thesis project. In order to answer my question about the temporal variation in human behavior, I would like to have dates that are much closer together and can identify patterns of sedimentation by century, at the very least since human movement and activity varies so sporadically through time. One aspect that may help to implement human activities with age-depth models is derived from the proposed question I left in the previous section associated with choosing large human occupational evidence as a hiatus in sedimentation. Regardless, this perspective may prove to be beyond the goals of my current research endeavors.
So, what does this current model (fig. 2) show? Within the parameters of the main age-depth function ‘Bacon’, I assigned the minimum depth at 0cm and maximum depth at 90cm. Furthermore, the minimum date was set to 13,000 cal BP and the maximum set to 16,500 cal BP. This was done because any date beyond these lies outside the archaeological context containing the current artifacts and dates. In the model, the light blue shapes indicate the statistically estimated dates that came from the 8 C14 samples (fig. 1). The shaded region around the red dashed line represents confidence intervals. It is clear that the more dates the model has, the tighter the confidence interval will be. This is evident by the expansion in the confidence interval ranging from 60-80cm. Ultimately, with an accumulation rate of 50yrs/cm, there appears to be a slight increase in deposition starting at a 25cm depth and a similar decrease in sediment deposition around the 55cm depth. Of course, this is just the surface of capabilities within the ‘rbacon’ package. From here, one would be able to identify deposition rate by year or by depth by producing graphics representing changes in depositional patterns (fig. 3). Figure 3 can be interpreted as the sediment accumulation rate per depth in centimetres and age in cal BP, respectively.
Ultimately, I am unsure how I will integrate this analysis into the interpretations of human behavior and occupancy at this archaeological site because, as I have noted previously, the main requirement for small scale studies requires a large amount of known dates for specific events or materials. That being said, the theory and understanding of how sediments may have accumulated throughout the time at this site is very important to consider when striving to make archaeological inferences and arguments. In the near future, I will begin to examine the temporal scope of artifact placement with sedimentation in mind. For this analysis, I will assume vertical displacement represents temporal displacement and analyze the rate of artifact deposition by depth to determine aggregates through time. Thus, allowing me to examine an observed temporal displacement of artifacts which differs from blog post 1 and 2 by adding the third dimensions, that is time.
Bailey, Geoff. “Time perspectives, palimpsests and the archaeology of time.” Journal of anthropological archaeology 26, no. 2 (2007): 198-223.
Blaauw, Maarten, J. Andrés Christen, K. D. Bennett, and Paula J. Reimer. “Double the dates and go for Bayes—Impacts of model choice, dating density and quality on chronologies.” Quaternary Science Reviews 188 (2018): 58-66.
Mandelbrot, Benoit. “How long is the coast of Britain? Statistical self-similarity and fractional dimension.” science 156, no. 3775 (1967): 636-638.
Plotnick, Roy E. “A fractal model for the distribution of stratigraphic hiatuses.” The Journal of Geology 94, no. 6 (1986): 885-890.
Sadler, P. M. “The influence of hiatuses on sediment accumulation rates.” In GeoResearch Forum, vol. 5, no. 1, pp. 15-40. 1999.
Schiffer, Michael B. “Formation processes of the archaeological record.” (1987).
As a reminder, my ultimate research goal is to quantitatively characterize salt marsh volumetric change from 1939 to the present within Oregon estuaries using a combination of sediment cores, for which I have excess 210Pb-derived sediment accumulation rates, and historical aerial photographs (HAPs), from which I plan to measure horizontal expansion and retreat. For my Exercise 1, I successfully georeferenced and digitized HAPs from Alsea Bay that are roughly decadal, spanning 1939 to 1991. The next major step was analyzing my data using the USGS Digital Shoreline Analysis System (DSAS; Himmelstoss et al. 2018). However, prior to analyzing my data in the DSAS framework, I had to first tie up a number of loose ends, which I did for my Exercise 2; these included determining an uncertainty associated with each shoreline layer, attempting to account for differences in shoreline resolutions related to different spectral resolutions of each aerial photograph, and determining modern shorelines (2001, 2009, and 2018). My next tasks included answering the following questions:
What final edits must be made to successfully run DSAS on my Alsea Bay shoreline layer?
What are the net rates of change in Alsea Bay salt marshes? (sub-question: from those that DSAS calculates, what are the most appropriate statistics to characterize rates of change?)
How does the Alsea Bay salt marsh shoreline change over time?
Step 1: Reduce unnecessary complexity in the shoreline layer
First, I removed all upslope edges on the fringing marsh since these are not of interest and all digitized using the same PMEP boundaries. I then removed all unnecessary complexity within the salt marsh complexes, especially within the northern portion of the larger island and the southern portion of the fringing marsh.
Step 2: Determine best DSAS settings using trial and error
Next I began to play around with the DSAS settings to produce the best results (assessed visually). These included:
transect search distance to remove transects bridging the gap between the largest and smallest islands. This may require running each marsh complex separately (which I would rather not due so that all transect symbology are on the same scale).
Play around with transect smoothing to reduce the number of intersecting transects in areas with complex shoreline.
Step 3: Calculate DSAS statistics
To calculate DSAS statistics, I selected the Calculate Rates tool, selected “select all” statistics, applied an intersection threshold of a minimum of 6 shorelines, and changed the confidence interval to 95%. DSAS reports the following statistics:
Net Shoreline Movement (NSM): NSM is the distance between the oldest and youngest shorelines
Shoreline Change Envelope (SCE): SCE is the greatest distance among all the shorelines that intersect a given transect, these values are therefore always positive
End Point Rate (EPR): EPR is the NSM value divided by the time elapsed between the oldest and youngest shoreline; even if the shoreline contracts, then expands so that the maximum distance (SCE) is not the same as the distance between the oldest and youngest layers, EPR and NSM values are still calculated based on the oldest and youngest layers
Linear Regression Rate (LRR): LRR is calculated by fitting a least squares regression line to all shoreline points for a transect; all transect data is used in LRR regardless of changes in shoreline trend or accuracy; this method is susceptible to outlier effects and tends to underestimate the rate of change relative to EPR. DSAS additionally provides the standard error estimate (LSE), the standard error of the slope given the selected confidence interval (LCI), and the R-squared value (LR2) for each LRR value
Weighted Linear Regression (WLR): WLR is very similar to LRR except it gives greater weight to transect data points for which the position uncertainty is smaller. DSAS additionally provides the standard error estimate (WSE), the standard error of the slope given the selected confidence interval (WCI), and the R-squared value (WR2) for each LRR value
Step 4: Rerun all previous steps with sets of dates from 1939 to 2018
Using all of the same steps DSAS, I reran each group of shoreline layers from 1939 to 2018. Following each run, I had to clip each rate layer using the DSAS Data visualization tool to fit the size of the transects specific to the timeframe of interest. I additionally changed the color ramps to be consistent for all 8 time points. LRR and WLR are not calculated for only two shoreline layers so EPR will be the primarly statistic I use for this time series analysis.
Step 5: Analyze DSAS statistics
I exported the DSAS rates files using the ArcMap Table to Excel tool. I then plotted and analyzed the data in MATLAB. I made histograms with Kernel smoothing function fits using the function histfit. I chose kernel smoothing over a normal distribution fit because the normal distribution fit did not effectively display the uneven tails or the humps on the sides of the NSM or EPR plots. I additionally made box and whisker plots of the EPRs over time for each marsh complex: fringing, large island, and medium island.
The histograms of the DSAS net shoreline movement for the entirety of Alsea bay approximates a normal, gaussian curve, with the majority of points falling the 0 m, indicating net erosion of the Alsea Bay salt marshes from 1939 to 2018. This could indicate widespread drowning of these marshes. These results confirm similar patterns of net vertical accretion rates lower than 20th relative sea level rise rates for the bay (Peck et al. 2020). Rates of growth (EPR, LRR, WLR) for the entire bay are additionally approximate a normal distribution and tend to be negative. LRR and WLR are likely better predictors of long-term rates of change since they better account for intermittent periods of erosion and contraction. Though the uncertainty does not vary much between shoreline year (~1 to 4 m), I will likely use WLR to ensure uncertainty is incorporated appropriately.
Preliminary results comparing the three salt marsh complexes display similar patterns as the total bay histograms. The large island values tend to display a greater range of movements and rates, especially towards growth. Despite these growth rates in the large island salt marsh, the majority of observations still skew towards erosion.
Preliminary analysis of the box plots shows that the EPRs have changed through time for each salt marsh complex. Box plots were likely not the best way to visualize this data so I plan to focus on this moving forward.
I have no negative critique of my methods so far for Exercise 3. I’m very excited with how the project is progressing. I think DSAS is a great method for addressing my questions. I’m still working on the best way to plot the figures, however. One next step I envision is creating a GIF displaying the colored transects that cycles through each shoreline layer from 1939 to 2018. Additionally, I would like to plot the mean EPRs for each time step for each salt marsh complex as bar plots with ranges. I think this will display trends better. I additionally think it’s time to start incorporating my salt marsh accretion data!
Himmelstoss, E.A., Henderson, R.E., Kratzmann, M.G., & Farris, A.S. (2018). Digital Shoreline Analysis System (DSAS) version 5.0 user guide: U.S. Geological Survey Open-File Report 2018–1179, 110 p., https://doi.org/10.3133/ ofr20181179.
Peck, E. K., Wheatcroft, R. A., & Brophy, L. S. (2020). Controls on sediment accretion and blue carbon burial in tidal saline wetlands: Insights from the Oregon coast, USA. Journal of Geophysical Research: Biogeosciences, 125(2), e2019JG005464.
Global Positioning Systems (GPS) technology has improved the ability of wildlife biologists to gain greater insight into wildlife behavior, movements and space-use. This is achieved by fitting wildlife, such as deer, with a collar equipped with a GPS unit. My research relies on GPS data representative of Columbian black-tailed deer (Odocoileus hemionus columbianus) in western Oregon to assess habitat and space use. Deer were captured and monitored from 2012-2017 in the Alsea, Trask, Dixon and Indigo Wildlife Management Units. Although this technology offers plentiful opportunities to answer research questions, GPS data is accompanied by errors in the form of missed fixes or inaccurate locations that should be corrected prior to conducting analyses. A common cause of GPS error is physical obstruction between the GPS collar and satellites such as topography or vegetation. One way to correct for habitat-induced-bias is to place GPS collars at fixed locations with known environmental characteristics and model GPS fix success rates (FSR) as a function of topographic and vegetative attributes. Nine GPS collars were left out for 5-7 day intervals at fixed locations (n=53) in the study area in February and March, 2020. Fix schedules matched the same fix schedule that was used on the deer collars. Locations were chosen specifically to represent a variety of terrain and vegetation in the study area. Results from the stationary collar test can be used to create a predictive map of FSR, and use as weights in subsequent habitat selection regression analyses.
1.1 What is the spatial extent of the influence of sky availability on GPS collar fix success?
1.2 What vegetative and topographic features best predict average GPS collar fix success rates?
Name of the tool or approach that you used.
2.1 I used focal mean statistics (a form of moving window analysis) to derive sky availability values at varying spatial scales. I performed focal mean statistics in R using the ‘raster’ package, and the ‘Extract Multi-Values to Points’ tool in ArcGIS to obtain values from test site locations. I used the ‘lme4’ and ‘MuMIn’ packages in R to perform linear regression modeling and model selection.
2.2 I used generalized linear mixed regression modeling to predict fix success rates as a function of environmental attributes. I used the ‘lme4’ and ‘MuMIn’ packages in R to perform linear regression modeling and model selection.
Table 1. Most explanatory variables for the stationary collar test were measured on the ground by technicians (canopy cover, slope, aspect, land cover type) but I created the sky availability variable (described in 3.1). The variables described here will be used in ‘Exercise 3’ to create a predictive fix success map across the study area.
Brief description of steps you followed to complete the analysis.
Response variable: Fix Success Rate
GPS fix data was downloaded from each of the 9 GPS collars, and FSR was calculated for each test site by taking the number of successfully obtained fixes by total number of attempted fixes. Collar data includes: coordinates, date, time, the number of satellites that were used to obtain a location and Dilution of Precision (DOP). DOP is calculated by the GPS collar, and is a measurement of accuracy based on satellite quantity and position. The higher the DOP, the less accurate the fix record is.
3.1 Moving Window Analysis
I loaded the sky availability raster into R and created new rasters with the ‘focal’ function. Six different window sizes were subjectively chosen based on what distances seemed reasonable to me to test: 30x30m, 50x50m, 110x110m, 210x210m, 510x510m, 1010x1010m. With this technique, each focal cell is reclassified as the average sky availability within each moving window.
After creating new raster data, I imported the raster stack of different scales into ArcGIS, along with the coordinates of the stationary collar test locations (n=53). I extracted the sky availability values using the ‘Extract Multi-Values to Points’ tool, and exported the attribute table as a .csv file to bring back into R.
A linear mixed model was used to predict FSR as a function of sky availability per test site, with collar ID as a random intercept because fix success can also vary between individual collars. I ran 7 univariate models and used corrected Akaike’s Information Criterion (AICc) to perform model selection. The top model was used in subsequent analyses.
Univariate generalized linear mixed model set that was used to compare FSR as a function of average sky availability at varying spatial moving window sizes.
3.2 Collar FSR Linear Regression
Most variables were measured on the ground by technicians (canopy cover, slope, aspect, land cover type) but sky availability was obtained as described above.
Using a generalized linear mixed model framework, I developed competing models using test site as the sample unit, and collar ID as the random intercept. Models were considered competitive if they were within 2 AICc units of the top model, and model selection was performed in a tiered approach. The first tier model set included univariate models of scaled covariates, interactions and any non-linear forms that seemed appropriate. Competitive models would proceed to subsequent modeling efforts.
Brief description of results you obtained.
4.1 Moving Window Analysis
Model selection output (Table 2) resulted in an ambiguous outcome. It appears that sky availability values within the moving window size of 110x110m is the magic number that best predicts variability in FSR, however the evidence for this is not very strong. All models appear competitive with delta AIC <2, and the top model only carrying 3% more of the weight than the next most competitive model.
Table 2. Model output from sky availability moving window analysis. All models were competitive (within 2 AIC units from the top model). The ‘top’ model was used in subsequent modeling for the purposes of this exercise.
4.2 Collar FSR Linear Regression
Judging from the raw data, the only nonlinear term that seemed appropriate was the quadratic form of canopy cover (Figure 1), and was the most competitive model with 73% of the weight in the model set (Table 3), and a conditional coefficient of determination of 0.72. The scaled covariate of canopy cover was the next most competitive model, but with a delta AICc score of slightly >2.
Table 3. Model selection results relating FSR to topographic and vegetative variables in western Oregon. The most competitive model was the quadratic form of canopy cover.
Critique of the method – what was useful, what was not?
5.1 Moving Window Analysis
My next steps are to re-examine the focal window type I used, and instead try an annulus style instead of the entire square to calculate the average sky availability values. This is in hopes that I will get a more clear result in which spatial scale is most important in sky availability influencing FSR. The ‘focal’ function was very easy and efficient.
5.2 Collar FSR Linear Regression
I was a little bit surprised that topography didn’t play a stronger role in influencing FSR, but canopy cover does make sense. However I am surprised that the quadratic form of canopy cover explained the most variation in FSR, because it doesn’t make sense that at lower canopy cover proportions FSR would also be low.
Moving forward, I plan to filter the fix data to reclassify locations that are likely to be inaccurate using cutoff values for number of satellites and DOP values. Doing so will decrease the overall FSR, and introduce more variability into the data and could potentially show different results.
Following exercise 1, (where I compared different interpolations to estimate zooplankton density), I wanted to make my interpolations more ecologically informed. I hypothesize that zooplankton will be found closer to kelp and that with increasing distance to kelp, there will be decreasing zooplankton density. Therefore, my question for exercise 2 was “Does zooplankton density decrease with increasing distance from kelp?”.
Approaches/tools used and methods
I used a simple linear regression in RStudio to address this question.
At the end of every field season, we use the theodolite to trace outlines of surface kelp. This produces many lat/lon points. Therefore, I first needed to create a kelp polygon layer from all of these points in R. Since I have never worked with spatial data in R before, it took me quite some time to figure out how to do this and to ensure that I was working in the correct coordinate system.
Next, I needed to calculate the distance between each of the sampling stations and the edge of each of the kelp polygons. Once this was done, I needed to select the kelp patch with the minimum distance for each of my 12 sampling stations.
Finally, I plotted the zooplankton densities for each of those stations against the minimum distance for each station, as well as ran a simple linear regression to see what the relationship between zooplankton density and distance to kelp is.
I was only able to run this analysis on 2017 and 2018 data because I am not yet finished with scoring the 2019 images for zooplankton density. However, the relationship is the same for both 2017 and 2018, namely, that zooplankton density decreases with increasing distance to kelp.
Since there is a big jump in distances to kelp (this is because two of the sampling stations are ‘controls’ which are over sandy bottom habitat where there is no kelp nearby), I also decided to log both axes to try and normalize the distance axis.
While I would have like to have explored a more complex tool or technique, this exercise has taught me to always keep your data in mind and to acknowledge the limitations that are contained within it. I only have 12 sampling stations and so am rather limited in the analysis that I can reliably run. Sometimes our data truly conform to a simple solution, which in my case, the simple linear regression seemed the most straightforward and obvious solution to answering my question about whether or not zooplankton density and distance to kelp are correlated. I feel pretty confident in the R2 and p-values that the above regressions show and so my next step will be to try and figure out how I can incorporate this linear function into my interpolations.
In Exercise 1, I tried to generate point data for the entire Oregon landslide. Due to the limitations of survey, the generated hot spot analysis does not show the situation of Oregon landslide. So in Exercise 2, I will first generate a hot spot analysis map of the relevant landslide volume, and select about 30 points of data through the hot spot analysis map to create a buffer zone. After that, the road density of the analysis area is calculated. By comparing the relationship between different buffer radius and road density, analyze which spatial scale are the relationships strongest.
2.Name of the tool or approach that you used
a) Hot Spot Analysis based on Volume of Landslide
Hot spot analysis is used, field is volume, Through the hot spot analysis map, find the gathering point, based on these points, select about 30 points to create a buffer zone as a reference object for comparison later.
b) Create a Buffer and Mutiple Ring Buffer
Buffer zone is created for these selected points, two different radius buffer are created, one is 250 meters, other is 500 meters. The reason to create buffer is to compare with road denstiy, to find strongest relationships in those spatial scale.
c) Line Density
In ArcGis Pro, the roads in analysis region are inport, and then using line density tool to calculate road density. As the reference for later comparison.
Figure 1 shows a heat map of landslides occurring in the city of Portland based on landslide volume. It can be found that the landslides in the mountains near Oregon city are similar in volume and cluster. Therefore, the buffer in Exercise 2 will be created around the surrounding point data.
Figure 2 shows the distribution of road density in the analysis area. The darker the color, the greater the road density. Through observation, landslides are roughly concentrated in the densest areas of roads. Table 1 shows the average value of road density in the area, which is 0.316.
Figures 3 and 4 show the relationship between roads and landslides in buffer zones with different radii. Although most of the landslides occurred around the mountains, it can be found that in the buffer zone of 250m and 500m, the greater the road density, the higher the impact. Analyzing Figure 3, Figure 4 and Figure 2 together, it is found that if there is a greater road density around the mountain, the more landslides that occur, the larger the affected area.
5. Critique of the method
In Exercise 2, my data is mainly concentrated around Oregon city in Portland, so when using hot spot analysis, I think it is efficient. The hot spot analysis based on the volume of landslides expresses whether the landslide data is clustered in Portland, and where is the most approximate data set. When using the buffer tool, due to the different data volume of different points, when creating a buffer zone, I am not sure whether to create a buffer zone based on its volume, or all data have a uniform radius.But for this exercise, this tool is effective.
With the data I had, I was interested in examining whether proximity to location of crop fires was associated with individual’s health outcomes. The reason that I had restricted my analysis to crop fires was because I want to explore whether there is a health-related argument for making illegal the burning of crop residue – which is the source of these fires.
Name of the tool or approach that you used.
I estimated the following logit regression model:
Hit = βo + β1EXPOSUREit + eit
where H is a binary variable for reported illness =1 if individual reported illness and 0 otherwise. Exposure is the count of fires experienced by the household in 5, 10 and 15 km distance respectively. It may be possible as a next step to control for other factors at a broader scale e.g. by including district level factors such as population and road density in the above regression model.
Brief description of steps you followed to complete the analysis.
In my data, I had point locations of daily fires as well addresses of household.
In order to better manage my data I assigned week of fire to my entire daily fire dataset. I then aggregated my data into 52 weekly fire files. Each fire had an associated longitude and latitude.
I was able to write out a code in stata that allowed me to calculate weekly counts of fires within a 5km radius. This resulted in 52 variables (weekly fires) with values corresponding to the count of fires in a buffer distance for each household.
I then summed up fires for the 10 weeks prior to the date of the interview to calculate the exposure variable for each household.
I repeated step (ii) and (iii) for a 10k and 15km distance.
As a last step before the regression I merged this with the individual level datafile.
Brief description of results you obtained.
Looking at the distribution of our sample (Table 1), it does not appear that any single demographic has a disproportionate exposure to fires. Across genders and all age cohorts 40% of individuals have had exposure to at least one fire within a 15km distance. About 15 % have had exposure to at least 1 fire within a 5km distance over the past 10 weeks. The count and percentage of individuals who have been exposed to fires within a 5km distance is highest in the province of Punjab. Given the distribution across provinces, it may be worthwhile to restrict the analysis (logit regression) to a sample based on Punjab and Sindh.
Up to 5 years
6 – 17 years
18 – 60 years
greater than 60
Table 1: Sample Distribution Note : The table presents count and percentage of individuals with at least 1 fire within a 5 km,10 km and 15 km buffer distance – in a 2-month period. Percentages are rounded to one decimal place.
The results of the logit analysis are presented in Table 2 below. The magnitude as a percentage of mean for two of the dependent variables – respiratory illness and blood pressure suggests a substantial positive association and confirms the original hypothesis. However, these are preliminary estimates and other explanatory variables will need to be included to gauge the impact of exposure to cropfires.
Table 2: Logit regression for the impact of exposure on reported illness
Mean = 0.013
Cough, cold, fever
Mean = 0.15
Mean = 0.013
*** p<0.01, ** p<0.05, * p<0.1. Standard errors reported in parenthesis.
Critique of the method – what was useful, what was not?
The logit regression is well suited for a binary dependent variable model and in this case, it is useful for estimating overall association in illness and exposure. It is also possible to predict errors from the regression which can be tested for spatial autocorrelation. If error terms are spatially correlated, it would suggest that the incidence of illness is spatially correlated. However, I was interested in exploring the temporal distribution of the fires, but I was not able to use the logit analysis for that purpose.
Exercise 2: Assessing the growth of artificially regenerated seedlings using UAV imagery
The question that is addressing in this study
Evaluation of tree health and growth after artificial regeneration is an important task, and monitoring of seedlings after the regeneration potentially provide vital attributes for the forest managers to make future decisions. However, monitoring can be unique to site characteristics and may complex based on site accessibility and terrain conditions. Hence the conventional monitoring approaches become more challenging. With the emergence of remote sensing technology, especially the applications of unmanned aerial vehicles (UAVs) provide a novel approach for seedling detection and monitoring, even in challenging terrains and remote areas (Buters et al., 2019).
After the artificial regeneration of seedlings, the growth and health of each seedling may depend on different parameters such as soil type, exposure to sun and wind, drainage conditions, space constraints, insect and disease susceptibility, etc (International Society of Arboriculture., 2011). UAV based monitoring approaches have high potential to classify and identify planets in disturbed forest environments and to monitor plant health as well (Berni et al., 2009; Lehmann et al., 2017). Thus, in this exercise, I am interested in developing an automated method to detect individual seedlings and their attributes (i.e., location, height, canopy area, number of trees in a given plot, etc.) from UAV based imagery (as the first step,). Additionally, in the second step of this study is focused on evaluating “How does the surrounding material affect the health of seedlings and its growth.”
2. Name of the tool or approach used, and steps followed.
Step 1: Detecting individual seedlings and their attributes
Canopy height model
I used R software packages to detect the seedlings (automatically) from the ortho-mosaic image.
Packages used: lidR, raster, sp, raster, rgdal, and EBImage
As the initial step, point cloud data were extracted using AgiSoft software (Fig.1(a)). Then the lasground function in lidR package was used to classify the points as ground or not ground by using the “cloth simulation filter” (Fig.1 (b)). The points which were classified as ‘ground’ are assigned a value of 2, according to las specifications (The American Society for Photogrammetry & Remote Sensing, 2011). In the next step, I used the lasnormalize function in lidR package to remove the topography from the point cloud created (Fig.1 (c)). Next, the normalized point clouds were cropped with the area of interest (AOI) shapefile (Fig. 2 (d)) using lasclip function.
Figure 1. Map of (a). point clouds generated over the study site, (b). point clouds after the cloth simulation function performed, (c). normalized point cloud, and (d). finalized point cloud mapped in AOI.
Then I used the output of the lasclip function to make the canopy height model, using grid_canopy command, and use 0.01m as the resolution, which is similar to the resolution of the raster image. Next, dsmtin algorithm was used to create the digital surface model. After that, a variable window filter algorithm was used to identify the treetops from a canopy height model. The minimum height for the detection of seedling was considered as 0.2 m for the canopy height model for a potential treetop. All canopy height pixels beneath this value will be masked out. The threshold of minimum height (0.2m) was used to minimize the effect of woody debris and small shrubs. Finally, tree seedling locations were exported to ArcMap and plotted them on top of the AOI raster to observe the performance of the algorithm.
Figure 2. Map of (a). area of interest (b). canopy height model, (c). locations detected of seedlings, and (d). canopy areas of detected seedlings.
Step 2: Geographically weighted regression
Image classification tools available in ArcMap 10.7:
Supervised image classification (SIL): Maximum Likelihood Algorithm (explained in blogpost 1)
Spatial Analyst Tool: Tools available in Zonal
Geographically weighted regression
Quantifying the surface materials
In order to identify the surface material surrounding each seedling, a supervised classification was performed. The supervised classification approach is more accurate and minimizes the effect of shadow and other noises compared to the unsupervised classification. The raster image was classified into five classes, including tree seedlings, woody materials, grass, barren land, and barren land.
As the next step, I created 3 types of buffer zones around each seedling with 1m, 2m, and 3m buffer radius (Fig.3 (a)). Next “Tabulate Area” option that is available in the “Spatial Analyst Tool” was used to extract the corresponding area for each class (Fig.3(b)).
Figure 3. Map of (a). 1-meter buffer zones around each tree seedling, and (b). a schematic diagram showing the extracted classes from the classified image by using a 1-meter buffer zone.
Geographically weighted regression
Geographically weighted regression is an extension of the traditional regression approach, which allows providing local estimation for each location instead of a sole regression for the entire area. In this regression, approach observations are weighted concerning their proximity to a point (point I), which is determined by the kernel size. “This ensures that the weighting of observation is no longer constant in the calibration, but instead varies with point I. As a result, observations closer to the point I have a stronger influence on the estimation of the parameters for location I.” (The Landscape Toolbox, 2012)
To determine the effect of surrounding surface materials on seedling growth, two types of geographically weighted regression analyses were performed using ArcMap 10.7. (1). Area of bare ground (Fig. 4) and (2). amount of woody debris (Fig.5) within three buffer zones (1m, 2m, and 3m) considered as the independent variable while estimated tree height considered as the dependent variable.
Effect of bare ground
Generally, bare ground indicates many indicators that are related to subsurface soil conditions such as insufficient nature of nutrients, the behavior of subsurface water table and water holding capacity. Hence, the presence of bare ground around the seedling can be used as an indirect indicator to predict the health conditions of seedlings.
Figure 4. Map of geographically weighted regression by considering seedling height as the dependent variable and area of bare ground within (a). 1-meter buffer, (b). 2-meter buffer, and (c). 3-meter buffer as the independent variable.
Effect of woody debris
The presence of woody debris around seedlings can cause both positive and negative impacts. For example presence of old tree stumps and debris act as microsites: especially when the seedlings shaded by stumps and logs tend to grow well particularly on hot, dry sites (Landis et al., 2010) However, the presence of woody debris/slashes may act as an obstacle to seedlings and limit the accessibility to sunlight and seedling growth. To study the effect of woody debris and their impacts on seedling growth I performed another geographically weighted regression by considering three types of buffers (Fig.5).
Figure 5. Map of geographically weighted regression by considering seedling height as the dependent variable and area of woody debris within (a). 1-meter buffer, (b). 2-meter buffer, and (c). 3-meter buffer as the independent variable.
Additionally, I created a map of normalized difference green/red normalized green, red difference index (NDGRI) to validate the results obtained by geographically weighted regressing (Fig.6 (a)). Furthermore, I generated a map of geographically weighted regression by considering the seedling height as the dependent variable and canopy area of seedling in the 1-m buffer zone as the independent variable by assuming healthy seedlings are taller and have larger canopy area.
Figure 6. Map of (a). normalized difference green/red normalized green, red difference index (NDGRI), and (b). geographically weighted regression by considering seedling height as the dependent variable and area of canopy cover within 1-meter buffer as the independent variable.
Bare ground coeffects:
Results obtained by geographically weighted regression using height and bare ground showed a decreasing trend of coefficient values with increasing the buffer size (Fig.7 (b)). Furthermore, with a given buffer zone, the values of the coefficients are not ranging drastically (only change can be observed in the fourth decimal place of the coefficients). These results suggest that the presence of bare ground within 1-meter and 2-meter buffer zones have a considerable impact on seedling growth. Obtained trends indicate that seedlings located in the northwest part of the study area have lower coeffects values, while seedlings located in the southeast part have relatively higher negative coefficient values (Fig.7(b)). A similar type of trend was observed for the area of woody debris as the independent variable: especially for woody debris within a 1-meter buffer zone, shows positive coefficients for seedlings that are located in the northwest part of the study region. Further, magnitudes of the coefficients are lower with increasing buffer size. Additionally, I evaluated the variation of coeffects (from both geographically weighted regressions) with seedling heights (Fig. A1, Appendix A). The observed results indicate that taller seedlings always tend to show either positive or lower negative coefficients.
Figure 7. (a). Map of seedling heights and corresponding seedling I.D.’s, plots of coefficients of geographically weighted regression by considering seedling height as the dependent variable, and (b). area of bare ground, and (c). area of woody debris within (1). 1-meter buffer, (2). 2-meter buffer, and (3). 3-meter buffer as independent variables. Trends of coefficients on the x-axis and seedling ID. on the y-axis.
By considering the observed results, we can interpret that area of the bare ground, and the presence of woody debris have an impact on the growth of seedlings. For example, trees growing surrounding more woody debris have the highest (negative) coeffect values and lowest heights. Similarly, areas that are prominent with bare ground show a negative effect on seedling growth. Results illustrated in figure 6 proves that trees growing in areas where bare ground and woody debris are prominent does not show a healthy seedling population. Overall, the results obtained from geographically weighted regression indicate the type and amount of material around the seedlings are vital, especially for their health and growth.
Possible background factors
Woody debris: generally, the presence of woody debris and their degraded products can provide nutrients to the soil and will help for the growth of seedlings (Orman et al., 2016). However, this process may vary with the size of the woody debris. For example, if debris are coarse, it will take a longer time to degraded and will act as obstacles for seedling growth. This type of observation can be observed in the southeast part of the study region (Fig. 2(a) and Fig.8 (a)). Furthermore, the northwest part of the region shows smaller size woody debris, and possibly they may degrade at a faster rate and provide some nutrients to the soil (Fig.2(a) and Fig.8(b)).
Additionally, down wood act as moisture reservoirs and provide persistence microsites for regeneration seedlings (Amaranthus et al. 1989). However, an excess amount of moisture conditions may negatively affect the seedling growth by rotting the root system of the seedlings. However, the presence of an excess amount of woody debris may tend to accumulate more moisture and will act as an adverse effect for seedling growth (as explained under Geography).
Figure 8. Seedlings surrounded by (a). coarse woody debris, and (b). small woody debris.
Bare Ground: Visual observation of bare ground can provide some essential facts about the subsurface soil conditions. For example, the bare ground can indicate several factors about subsurface soil conditions(i.e., lack of nutrients, problems associated with groundwater table, soil pH, subsurface geology, etc.). Hence, it is essential to have some prior information about the subsurface soil conditions to understand possible factors that are affecting seedling growth. Overall, we can interpret that seedling growth is lesser with the presence of bare ground; therefore, we can interpret that poor soil conditions hinder seedling growth.
Geography: To identify the possible effects of terrain conditions, especially for the results obtained from geographically weighted regression, I generated a digital elevation model (Fig. 9) and searched for possible prior information on site conditions. The digital elevation model suggests that the majority of healthy seedlings located in the northwest part of the study area with relatively high elevations (Fig.2(a) and Fig.9). Contrastingly, seedlings growing under stressed conditions in the southeast part of the study region were located at lower elevated areas (Fig.2(a) and Fig.9). Consequently, we can hypothesize that the accumulation of the excess amount of water (groundwater or surface water) in lower elevated areas can negatively affect seedling growth. This might be the indirect cause for the presence of bare ground in lower elevated areas (i.e., stagnant water may damage the root systems of grass and may die due to rotten roots and will expose the bare ground). Similarly, the presence of excess water may hinder the growth of seedlings as well. As a side effect of the presence of groundwater with woody debris may tend to increase excess moisture conditions in subsurface soil layer and will act as a negative factor for seedling growth, especially for their root system.
Figure 9. Digital elevation model of the study area.
Overall, the automated detection of seedlings and derived attributes show promising results. However, the presence of woody debris may affect the detection of trees, especially for small unhealthy seedlings. For example, the developed model will falsely detect piles of woody debris as seedlings. According to the results obtained from the canopy height model, the heights of most of the woody debris are less than 0.2 m. Hence, to avoid this type of interference, I introduced a threshold value (0.2 m) for the canopy height model.
Additionally, lack of prior background information (i.e., ground conditions, soil and hydrological conditions, quantitative information of woody debris, etc.) became quite challenging when interpreting the results of this study. Therefore, the collection of in-situ information would be beneficial for future research activities. Additionally, prior assessment of woody debris and their decaying nature would be valuable as a supportive attribute for estimating seedling health.
Figure A1. Estimated coefficients from geographically weighted regression for the area of bare ground (top) and area of woody debris (bottom) with seedlings height.
Amaranthus, M.P.; Parrish, D.S.; Perry, D.A. 1989. Decaying logs as moisture reservoirs after drought and wildfire. In: Alexander, E., ed. Proceedings of the watershed 1989 symposium on stewardship of soil, air, and water resources. R10-MB-77. Juneau, AK: U.S. Department of Agriculture, Forest Service, Alaska Region: 191–194.
Berni, J.A.J., Zarco-Tejada, P.J., Sepulcre-Cantó, G., Fereres, E., Villalobos, F., 2009. Mapping canopy conductance and CWSI in olive orchards using high resolution thermal remote sensing imagery. Remote Sens. Environ. 113, 2380–2388. https://doi.org/10.1016/j.rse.2009.06.018
Buters, T., Belton, D., Cross, A., 2019. Seed and Seedling Detection Using Unmanned Aerial Vehicles and Automated Image Classification in the Monitoring of Ecological Recovery. Drones 3, 53. https://doi.org/10.3390/drones3030053
International Society of Arboriculture., 2011. Tree Selection and Placement [WWW Document]. Tree Sel. Place. URL https://www.treesaregood.org/portals/0/docs/treecare/TreeSelection.pdf
Landis, T.D., Dumroese, R.K., Haase, D.L., 2010. Seedling Processing, Storage, and Outplanting, in: The Container Tree Nursery Manual.
Lehmann, J.R.K., Prinz, T., Ziller, S.R., Thiele, J., Heringer, G., Meira-Neto, J.A.A., Buttschardt, T.K., 2017. Open-Source Processing and Analysis of Aerial Imagery Acquired with a Low-Cost Unmanned Aerial System to Support Invasive Plant Management. Front. Environ. Sci. 5. https://doi.org/10.3389/fenvs.2017.00044
Orman, O., Adamus, M., Szewczyk, J., 2016. Regeneration processes on coarse woody debris in mixed forests: Do tree germinants and seedlings have species-specific responses when grown on coarse woody debris? | Request PDF [WWW Document]. ResearchGate. URL https://www.researchgate.net/publication/304404774_Regeneration_processes_on_coarse_woody_debris_in_mixed_forests_Do_tree_germinants_and_seedlings_have_species-specific_responses_when_grown_on_coarse_woody_debris (accessed 5.8.20).
The American Society for Photogrammetry & Remote Sensing, 2011. LAS SPECIFICATION VERSION 1.4 – R13.
How the spatial and temporal correlation between the Particulate Organic Phosphorous (POP) and PP (Primary Production) in Bermuda Ocean.
My raw data is unique, where the dates and coordinates of POP and PP is not matching each other, there are only eight data which match, hence I need to adjust it into several ways:
For spatial correlation, I expected that the nearest coordinate of POP to PP will have a very high correlation. In my hypothesis, Primary production is affected by the Phosphorus which came from the dust deposition. The location of this sampling is on the southeast ocean of Bermuda island, hence it is expected that the dust deposition is originated from the Sahara desert.
For temporal correlation, I divided the correlation based on the four seasons. Since the site is in the northern hemisphere, hence winter is from December to February, spring from March until May, summer from June until August, and autumn from September to November. The dust from Sahara will be transported in summer, where the heavier particle will deposition first near to the Sahara Desert and the small particle will deposition latter by the precipitation. Thus, based on this, the hypothesis is there will be a high correlation between the POP and PP in autumn and winter.
Tool and Approach
At first, I would like to use the Neighborhood analysis, however after I accessed the raw data thus, I divided the data based on the nearest distance between POP and PP.
ArcMap 10.7.1. I used join table for POP and PP data which based on the date since the coordinate is different and the FID has a different length.
MATLAB. I used MATLAB to average the data based on distance and generate the R square and scatter plot.
Description of Analysis
The spatial category is based on the distance between POP and PP wherein this step I used the Pythagorean formula. Subsequently, I divided the data based on three categories: I (0-0.030 degree), II (0.031 – 0.67 degree), and III (0.68 – 0.9 degree).
At first, the temporal category is based on all data from 2012 – 2016, and then I also analyzed the temporal pattern based on the distance category.
Time Series plot to see the propagation of POP and PP over five time period.
Scatter plot and Regression analysis
A Scatter plot is done for both spatial and temporal to see the correlation between the POP and PP spatially and temporally. The regression analysis is done to prove the significance of the correlation.
w = regress(POPw,PPw)
tblw = table(POPw,PPw);
lmw = fitlm(tblw,’linear’)
Time Series POP
From figure 1, we can see there is a lag in temporal patter, wherein 2012, there is no record on may, while in 2014 there is no record from January to February and in 2015 there is no record for January and October. In addition, in 2016, the data for winter is only available in March. However, overall we can see that the POP pattern is clustered from January to August, especially in 2013-2015, where the highest POP is converged in January and April.
2. Time Series PP
From figure 2, the lag also appears, wherein 2013, there is no record in March, while in 2014 there is no record in January, February, and May. In addition, in 2015 the temporal lag is similar to POP and in 2016 there is no record from January to February. However, almost similar to the POP, the temporal pattern is clustered from January to July, so does with the result, the highest PP is concentrated from January to April.
3. Scatter Plot based on Seasonal
From figure 3 we can conclude that the high concentration of PP happens in winter which until 20.87 mgc/m3/day which also followed by the PP which until 0.034 umol/kg. However, even the PP in spring is also high, however, we can see that the value of POP is concentrated on 0-0.01 wherein this value, the PP is relatively high. However, after I tried to prove this finding by matching it to the R square, it is found that the highest R-square is on autumn (0.17) compare to another season, where the lowest is on spring (0.00458), followed by winter (0.0773) and summer (0.0846). It is indicated that the deposition of POP in the Bermuda ocean is based on a dry and wet deposition.
4. Scatter Plot based on distance category
Figure 4 is the nearest distance measurement between POP and PP and the result is not likely what I expected, where in autumn the PP is relatively high, however, unfortunately, it is not followed with the POP value, where the lowest POP value happens on autumn. The R square indicates that in this category, the summer season has the highest correlation (0.227), followed by spring (0.117), autumn (0.00146), and winter (0.000297).
As overall, figure 5 indicates the lowest R score compare to category distance I and III in figure 4 and figure 6, which about 0.0036. However, if we see per seasonal, hence winter has the highest R score (0.0613) and summer is the lowest (0.0117). However, there is no significant measurement on winter compare to other season. If refer to the hypothesis, this result of this category matches the hypothesis very well.
From figure 6 we can see the relationship of the farthest distance between POP and PP. The result of R square is similar to figure 5, where the highest is in winter (0.509). However, the trend value of PP is different to category I and II, where the highest PP score for category III is in on summer, whereas two others are in autumn.
Critique of the method
My raw data is a little bit messy, I need to think more about how to process it properly. I am not satisfied with the analysis and the result that I got this my exercise 2. I did not access the normal distribution before doing the regression, which I guess makes me cannot get a significant finding and the R-square is very low compare to the global average. However, by dividing the data into several categories make me can determine the spatial and temporal pattern that I can clip to the Hysplit back trajectory data because in that step I need exact coordinates and time to deploy the modeling.
My question for exercise 2 involves the relationship between wood cover (in this case, wood area (m^2)/ plot area (m^2)) and channel width through plot center (m). My response variable is wood cover, and my explanatory variable is channel width.
I used geographic weighted regression (ArcGIS Pro) and simple linear regression (R Studio) to assess this relationship.
My initial approach involved using the wood accumulation index (WAI) raster I developed in exercise 1. It became clear this was not the best approach, and Julia and I agreed I should attempt to investigate the relationship between wood cover and stream width. My hypothesis was as channel width increased, wood should accumulate and therefore we will see a higher wood cover in plots at wider portions of the channel.
Step 1. I delineated a length of stream that approximated the main flow of the South Fork McKenzie River (Fig. 1). This line feature was approximately 850m in length. In order to produce enough observations for regression analysis, I divide length by 30, and plotted points even spaced along this line at 28.3m intervals.
Step 2. Next, I produced plot around these points with 10m buffers and used calculate geometry tool to calculate the overall area encompassed by these plots 312 m^2).
Step 3. I dissolved all features classified as wood into one feature. I then intersected this shapefile with my plots shapefile. This resulted in 30 in-plot wood features. Next, I made a new column for wood area and calculated area for these 30 features.
Step 4. To develop a wood cover variable, I created another column and used the field calculator tool to calculate wood cover: (WoodArea) / (PlotArea) * 100.
Step 5. To produce my explanatory variable (channel width through plot center), I used the measure tool to measure a line segment from one side of the channel to the other that passed through plot center. I approximated the shortest segment, and I recorded these observations in a new column (StreamWid).
Step 6. I used the geographic weighted regression tool to calculate coefficients for StreamWid. These were all very low
Step 7. I imported the associated database file into R studio and performed a simple linear regression.
Results from this exercise are not promising. The p-value associated with my linear regression model (figure 2) was 0.72. Therefore, including stream width in the model seems to be insignificant. Additionally, the associated r2 is .005 indicating very little correlation between stream width and wood cover.
Figure 2: Simple Linear Regression Model
Results from the GWR were similar. Local r^2 values were low and ranged from .009 to .016 (figure 3).
Figure 3: R^2 results from GWR
The approach I used for this exercise is sound and easily applicable to other hydrological modeling (albeit, this is not my realm of expertise). Stream width does not appear to correlate with wood cover in the particular flow transect that I sampled. However, it may be interesting to explore other explanator variables including distance to nearest land feature, water indices, or various plot sizes.
As the title of this section alludes to, I chose to assess the spatial interactions that categorical artifacts have with each other in space. The assumption that must be made states that areas of the archaeological that once contained human activities creates a source for specific artifact patterns. This assumption can then be broken down into a series of additional corollaries:
C1: Human behavior associated with domestic activities will result in a source for a medium count of diverse artifact types (i.e. charcoal, bone, debitage).
C2: Human behavior associated with tool production will result in a source for formal and informal tools, a high abundance of debitage, no abundance of bone and charcoal.
C3: Human behavior associated with animal processing will result in a source for blades, scrappers/uniface, utilized flakes, medium abundance of debitage, high abundance of bone, and no abundance of charcoal.
C4: Human behaviors that produce features such as hearths and pits create a strong source for a diversity of artifacts within an immediate vicinity of the feature.
The above corollaries walk through an assumption that is generally made about archaeological remains when observing the distribution of artifacts. Each on of these corollaries can be broken down even further and more specifically but, because I am not conducting an experimental archaeological investigation (in which these corollaries would be hypotheses), I have chosen to maintain a presence of generality. Additionally, in my statistical calculations, I have only included the three most abundant artifact types to analyze, debitage, bone, and charcoal. These provide for adequate sample sizes and I hypothesize that they share the most amount of variance with human activities in the past.
The questions I wish to address for this blog post is: How does debitage, bone, and charcoal interact with each other to form clustered or dispersed boundaries reflecting human use and perception of space throughout the site, in one point in time and through time?
Cautious Approach to Spatial Autocorrelation
To assess the question stated above, I used spatial autocorrelation on debitage, bone, and charcoal at this archaeological site. Using Ripley’s K and the Pair Correlation Function (PCF) through R, I analyzed the autocorrelation between all of the artifacts, individual artifacts, and across each artifact type at the horizontal intervals of 0-25cm, 25-50cm, 50-60cm, 60-74cm, 74-100cm. I judgmentally chose these segments through the steps outlines in my blog post 1. By using autocorrelation, I hope to identify boundaries, clusters, and a dispersion of human activities across the site at each interval. This type of analysis must be taken with caution because, as an archaeologist, I cannot assume that 10-25cm intervals contain a single occupancy period at this site. Therefore, by assessing these intervals, I am not defining a single point in time, but I am arguing that these intervals indicate shifts in how people at the site organized themselves in terms of human behavior and spatial perception.
An additional cautionary method I needed to implement was the correction for inhomogeneity in my dataset by using ‘Kinhom’ for Ripley’s K and ‘pcfinhom’ for the PCF in the spatstat package. This is a necessary step to undergo if the data being analyzed is not homogenously distributed across the study area. This aspect of data can be easily assessed through the production of density maps showing observable clusters. Ultimately, the inhomogeneous correction for these autocorrelation methods more realistically addresses the observed and Poisson distributions.
Now, onto the fun part.
Figure 1 shows the results of Ripley’s K and the PCF operated on an interval of 0-25cm. I had to segregate the boundary of the archaeological site due to the excavation methods and amorphous topography of the surface. Therefore, I am under the assumption that the archaeological data here is an adequate representative of human behavior within the whole site. By comparing Ripley’s K and the PCF, there is a suggested a tendency for clustering (above the red line) from 0-20cm (not significant) and from 90-115cm (significant) at this interval. From the density map, it is clear that the clustered areas reflect the size of 90-115cm but there is also evidence for suggested clustering from 0-20cm which may not have been highlighted in the density map.
Figure 2 shows Ripley’s K and the PCF operated on the interval of 25-50cm. Ripley’s K and the PCF identify clustering at 0-15cm (significant). Additionally, the PCF shows clustering at 100-120cm. These are fascinating results because from looking at the raw data one might infer a larger cluster located near the left margins of the site, but the functions both reflect two sized clusters that are identified through the density map.
Figure 3 shows Ripley’s K and the PCF operated on the interval 50-60cm. In this instance, the results suggest a rapid drop of in clustering within close proximity and clustering starting at 15-25cm (mot significant). This best represents the spatial location of artifacts within the raw data map because the density map appears to reflect slightly larger clusters.
Figure 4 is my personal favorite because of the fascinating oscillation that occurs around the boundary for random distributions. Ripley’s K and the PCF both show significant clustering at 0-20cm. Any artifact past 20cm appears to be randomly distributed within space. Again, this slightly contradicts the observable characteristics in the raw data and density maps, leading to further investigation when imposing specific clusters and human behavior within this surface in the future.
Figure 5 shows a tendency for clustering at 0-15cm (not significant) with two further tendencies for clusters at 25-40cm (not significant) and 60-80cm (not significant). These cluster extents can be seen within the raw data and density maps but may warrant further analysis prior to assigning a specified number of clusters to this segment.
Lastly, figure 6 shows the cross correlation between each artifact type within a PCF corrected for inhomogeneity compared between the 50-60cm interval and 60-74cm interval. This figure helps relate to my initial question as to how each artifact relates to the other within space. By addressing shifts in artifact-artifact relationships, I hope to strengthen my argument that people between these two intervals perceived and used space differently at this site. Through these graphics, there are some similarities and differences that we can identify. The correlation between bone and debitage is suggested to be closer at 60-74cm, while it begins around 20cm (not significant) at 50-60cm. The correlation between charcoal and debitage is suggested to be fairly similar between the intervals. This aligns with my corollaries showing that debitage and charcoal interact with one another only in specific domestic human behaviors. So, it can be assumed that this specific interaction may influence similar spatial patterns through time. The last correlation between charcoal and bone at these intervals is fascinating and warrants further investigation because the graphics almost appear to be mirror reflections of one another, showing a closer correlation at the 50-60cm interval.
The major critiques I have for this approach is the affect borders have on statistical interpretations and, in a similar vein, the scale of data. Borders have always been a barrier within statistical analyses (pun intended). The problem with borders is the absence of data that lies beyond it that is both there and not there (kind of like Schrodinger’s cat). I would have to do more research into the specific mathematical properties affecting by borders, but it is important to understand that they do influence both calculations and interpretations. The second critique on scale is also important because Ripley’s K and a PCF can be operated when there are very few data points, but it will always suggest one’s data is always dispersed. From a mathematical view, this is a correct and appropriate assumption but is trivial to interpret by simply looking at your raw data. So, do not waste time with trying these functions on sparse data that is visibly dispersed already.
For Exercise 2 I changed the dataset and question that I am asking, so I will be sharing the results from a hotspot analysis (which would have been exercise 1) and a geographically weighted regression. I am interested in identifying gentrification in Portland, OR so I am looking at race, income, and education data at the census tract level. I got these data for 1990, 2000, and 2010, and have been experimenting to see which will be useful. An important note is that census tracts change shapes over the years, so for this class I am using the 2010 census tracts and acknowledging that there is an error associated with this decision. I obtained the census tracts for all of Multnomah County and trimmed it down to exclude large census tracts on the outskirts with small populations, census tracts with no data in 1990, and some of them east of the city of Portland (closer to Gresham).
The race and education data came as counts and the income is median income per census tract. I added in the raw population numbers in order to calculate race and education percentages. These are percentage of population that is white and percentage that has a bachelor’s degree or higher.
The raw percentages of white population show that in 1990, overall Portland is over 80% white in most tracts except for between the Willamette River and the Portland Airport. Over the next decade the who eastern side of the city becomes less predominantly white and continues to 2010. The area in the northeast that was less white to begin with had an increase in proportion of white population.
I conducted a hotspot analysis of the changes in each demographic from 1990-2000 and 2000-2010. This turned out to be very useful for finding the patterns in the data (more useful than when I ran it for my previous data). Below I have shown the hotspot map for change in race from 2000-2010 (which is similar to that of the previous decade). It becomes very obvious that there is an increase in white people (in red) in the north part of Portland (between the airport and the Willamette river) and a decrease in the eastern part of the city. I found this visualization of race data made by Portland State University that displayed a dot per person and assigned them a color indicating their race. The map swipes between 2000 and 2010 and confirms the pattern I see but provides insight that white people are replacing the black population in the northern part of the city and the population of people with Asian descent is increasing in the east.
The hotspots for both education and income are generally in the same regions as the hotspots for race, in both decades, but are smaller. This led me to the next step, which was conducting a geographically weighted regression.
Understanding the spatial distribution of the different variables and their correlation is important to understanding their relationship. Since my thesis research will entail mapping gentrification in order to map the spatial correlation with urban gardens, I need to ensure that I have a reliable source of information for gentrification. My hope was that I could use these three variables to map gentrification. My hypothesis for the pattern between the three variables is that they will be relatively similar. I imagine in such a white city there will be places with increases in income and no change in race. I also understand that there is an increase in Asian population in the east, which unlike black populations is not necessarily lower income on average.
I conducted the Geographically Weighted Regression with the ArcGIS tool a variety of ways. I tested each of the three different variables as the dependent and explanatory variables, and tried with just one explanatory variable, and two explanatory variables. Overall, I think this tool did not provide extremely useful results because there are too many assumptions that are not valid. The results show that the variables are somewhat correlated. The standardized residual maps show that most of the census tracts have similar changes in the three variables. When change in race is the dependent variable, and education and income are both explanatory variables, the northern region that is likely gentrifying is identified as having a higher standardized residual. The graphs show that change in race and change in education have some correlation but change in income has very little correlation with either race or education. The graphs below show these relationships between the variables in 2000-2010.
Overall, I think that the Geographically Weighted Regression tool was not extremely useful. It helped identify that there is some correlation between the change in race and education, but the hotspots show a slightly different story. The hotspots show that the three variables have increases in similar parts of the city. It is possible that I am not using the tool correctly, but I believe that the assumptions needed to utilize the GWR tool are ones that I cannot make with this data.
My goal for this exercise: to identify to what extent my county-level data is spatially and temporally autocorrelated. This is part of a larger effort to identify a suitable data sample that does not violate the assumption of independent observations.
Exercise 2 significance in context: Do we see significant out-migration in instances of high flood claims or is there simply a trend of adjacent counties having similar values due to space and time similarities? Or simply, is this a local or larger scale phenomenon?
2. Tool or Approach Utilized
While originally I was interested in attempting to apply geographic weighted regression to examine the spatial relationship of flood insurance claims and migration flows. After reading in-depth about GWR I found that not all assumptions of Ordinary Least Squares could be satisfied. Thus, I was interested in finding suitable subsets within my data that could be used for analysis. I was primarily interested in spatial and temporal autocorrelation (the assumption that all observations are independent). To examine the extent of autocorrelation I created subsets of my data and created a series of scatterplots to visualize spatial and temporal relationships of adjacent counties.
3. Description of Methods
I first created 2 subset datasets, one of urban well-populated adjacent counties and the other of rural low-populated adjacent counties all in Florida. Florida was utilized due to its high participation in flood insurance. I used Broward county as the metro county of comparison (High population center) to compare with adjacent high population counties. I used a subset of 3 rural low population counties in Northeastern Florida where Dixie County was selected as the primary county for comparison with adjacent rural low population counties.
I then explored trends in the data related to spatial autocorrelation by urban and rural out-flow rates over time. I then plotted 1 urban county out-flow rate against adjacent counties out-flow rates and did the same for rural counties. I then did the same thing for a hurricane period of time and a non-hurricane period of time. Finally I followed the same steps for hurricane and non-hurricane periods but with the variables or in-flow rates and claims per county population rates.
Next, I examined temporal autocorrelation by selecting a county and plotting out-flow rates at time 1 on 1 axis and time 2 on another axis for both hurricane and non-hurricane periods with both urban and rural counties. I then tried this with the variable of claims per county population.
4. Brief Description of Results
I first examined the general trend of out-flow rates across the whole time period (1990-2010). Looking at the above plots above we can see there is a general relationship between out-flow rates in spatially adjacent counties across time.
A) Hurricane Period Out-Migration Plots
Looking at these 2 plots there are obviously far fewer data points meaning they will be much more affected by outliers. From the metro hurricane period (1998-2001) we see similar out-flow rates with higher rates in adjacent counties for this period. For the rural hurricane period (1994-1996) we see generally higher out-migration rates with a somewhat positive linear relationship. Meaning that as Dixie County out-migration rate increases the adjacent county out-migration rates also increase (not causal).
B) Non-Hurricane Period Out-Migration Plots
The metro county out-migration rates reflected here are for (1990,1993,2002, 2003) the lowest insurance claim years and the rural county years are (1999-2002). We see no real discernible trend for non-hurricane metro out-flow rates which seem approximately similar to the metro hurricane years plot above. The rural non-hurricane period also does not demonstrate a significant linear trend with approximately the same values as hurricane years. This likely due to the small number of data points. I saw similar results when plotting the variables of in-flow rates and claims per county population rate.
Temporal Autocorrelation Plots
A) Hurricane Period Time 1 and Time 2 for single County Out-migration Rates
These plots are the result of the subsetting of Broward County (metro) and Wakulla County (rural) out flow rates at different times for the X and Y axes during a Hurricane period. We can see that out-flow rates seem relatively constant across the different times indicating evidence of temporal autocorrelation.
B) Non-Hurricane Period Time 1 and Time 2 for single County Out-Migration Rates
These plots give similar results to the hurricane period however the Broward plot does not have enough non-hurricane time periods to show this trend. As indicated for the hurricane time period we see evidence of temporal autocorrelation here. Similar results were seen with other variables such as claims per county population and in-flow rates.
Critique of Method
This was a useful exercise to help examine the spatial and temporal relationships at a small scale. I would say this was useful in painting a partial picture of the extent of the autocorrelation issue, however, not a complete one. This method did not help me narrow down a sample dataset to use for future methods and I am unsure of what my next steps are in continuing with this data in an attempt to fit an econometric model. I would be curious to find some sort of model that allows looser assumption requirements or transformation methods that may help me overcome these autocorrelation issues.
I decided to simplify my exercise 2 for a number of reasons, primarily because raw annual data turned out to be more interesting than I anticipated, and because I wanted to leave some tools/questions behind for exercise 3 (also, generating annual data was not a cake walk). As a result, exercise 2 has become entirely focused on the annual correlation between terminus retreat in a glacier, and the corresponding glacier velocity at the terminus location. ArcGIS pro was used to generate the terminus distance retreats for each available year, to create very small buffers around termini, and to extract underlying velocity raster values. Once a master-dataset was generated, I assessed their relationships in excel.
1.For any given year in my dataset, what is the correlation/relationship between the distance of individual glacier terminus retreat, and individual glacier velocity? (on individual scale and Greenland wide scale)
Simplified: At “T1” how is local glacier velocity correlated to local glacier retreat?
Context from our lab exercises:
What kinds of enhancing or limiting processes might occur that allow B to influence A, in space?
I presume that glacier velocity and retreat distance will be linked via the mechanism of basal melting (wet basal conditions will not only allow glaciers to move faster, but can also be representative of negative mass balance). Fully understanding basal slip is still a major problem in glacier physics. My hypothesis is based purely on the fact that glacier velocity seems to increase early in the melt season- indicating it is linked to negative seasonal mass balance. The spatial influence is obvious in that the retreat and velocity are being measured on the same glacier. More distal influences could include snow accumulation and ice temperature.
What kinds of enhancing or limiting processes might occur that allow B to influence A, in time?
This is a more difficult question. I would think that a glacier that experiences basal slip at T would be more likely to experience it at T+1, in a sort of positive feedback. Basal melt could reduce mass balance, which reduces thermal inertia and in turn results in more basal melt. That however does not explain how a brief pulse of velocity at T, followed by a slowdown at T+1, would influence terminus retreat at T+1. I cannot think of a mechanism in which velocity at T would influence terminus retreat at T+1 in this scenario.
Buffer, Construct Pyramids, Extract Multivalues to Points, Join field, Table to Excel
Excel Tools used:
Time series (plots), R-squared regression analysis, normalization
Data was manipulated as such:
(step 0: Using “Approach #2” from exercise 1, annual datasets of terminus retreat were generated, and annual velocity rasters were added to the map)
1.The BUFFER tool was used to create very small 500m buffers around glacier point data- (because velocity data comes at 500m resolution, this can be thought of almost like a point source- but not quite)
2.For some reason the raster datasets had pyramids, so I had to go through and use CONSTRUCT PYRAMIDS and set each to 0.
3.The EXTRACT MULTIVALUES TO POINTS was used to extract and average the velocity values sitting directly beneath buffer shapes, for each year in the dataset. This data is automatically added to the buffered-shapefile feature class attribute table. (see below)
4.The TABLE TO EXCEL tool was used to export the buffer attribute table, containing terminus retreat and underlying velocity data for every available year, to an excel file. I find it easier to manipulate and visualize data in other software, as Arc can get a bit clunky.
5.Time series plots containing both distance retreated and local velocity were constructed for every glacier and visually inspected.
6.R-squared linear regression plots were plotted together for each year to assess statistical correlation between the two variables at T1
7.Glacier velocities and retreats were normalized on individual glacier scales (each year subtracted mean and divided by it), to eliminate absolute magnitude effects.
8. Normalized data was plotted in R-squared linear regression plots.
9. Time series were averaged over the entire island on an annual basis, and plotted, visually inspected.
The methods used ended up being simpler than I anticipated, because the results I ran into were interesting and perplexing, and I wanted to investigate them further. To assess glaciers on an individual scale, I created a map with inlayed data (I originally intended to include more time series but I ran into an error with data alignment that I need to figure out, hence some arrows leading to nowhere) :
Because I assume this will come out at terrible resolution in the blog post- in the time series: Red lines represent the glacier velocity and blue lines represent terminus retreat- plotted by year. The raster overlay is Greenland surface velocity, in the year 2017, measured in meters/year. The points represent each glacier terminus and are colored according to their overall retreat (2005-2017) For some glaciers, velocity over the years seems to roughly follow the retreat particularly in the Northwest of the island. For other glaciers, notably several on the East coast, the glacier velocity does not appear particularly related to terminus retreat. Another purely visual assessment is that, on identical color scales, the dots (retreat magnitudes) do not always seem to match up with underlying velocities. In simpler terms, just by the eye test, the fastest glaciers don’t always necessarily retreat the fastest. But perhaps there is an overall trend? To assess, I made a simple r-squared linear interpolation graph for each of these (~240) pairs, color coding each point by year:
The results indicate no clear correlation. Each point on the graph represents an individual glacier in a specific year, so each glacier is represented a total of 7 times. The linear R-squared relationship between these 2 variables is never greater than .002, for any individual year. As a result we can say that there is no correlation between ABSOLUTE magnitude of glacier velocity and ABSOLUTE magnitude of glacier retreat. That still doesn’t explain why a fair number of glaciers appear to have curves that follow each-other on relative scales. After all, when each year is averaged out among all 240 glaciers, we get the following time series graph:
Here, we notice two things in particular. First, the overall curves seem to covary to a certain extent. Second, there appears to be two “steady states” so to speak: one pre-2008 and one post-2008. Before 2008, Ice retreat and ice velocity were both almost negligible compared to their post-2008 counterparts. The correlation between the two can be isolated using 240 glacier averages over the course of the study period plotted together:
Here, each point represents a single year. When annually averaged across all glaciers, the r-squared of termination retreat vs. ice velocity becomes 0.53. When 2006 and 2007 are removed, that R-squared jumps to 0.65. These values are more in line with what I was expecting to see based on the map and time series produced above. So how can I reconcile the two linear regressions? Clearly the two variables are connected in some way, but I need to figure out how. My first thought is that perhaps I need to eliminate the absolute magnitude of the variables in each glacier, and normalize them into relative magnitudes. That way I can compare relatively large years and small years in a less biased way aka: (for each glacier: does a relative change in velocity correlate to a relative change in terminus retreat?) I normalized measurements for each glacier by subtracting annual measurements from the study period mean, and dividing by the mean, plotted here:
While the R-squared is no better here not even displayed, the graph does depict a reality with two clear steady states- one pre-2008 and one post 2008. Normalized velocities were miniscule before 2008, and then quite significant after. Furthermore, the normalized terminus retreat appears to be far more variable in the years post 2008 (they tend to span a taller height on the graph). These are clear trends I want to dig further into in exercise 3.
If I were to go back and do this exercise over again, I would probably increase the size of the buffer. In creating such small buffers, I created some shapes that were not even overlapping with velocity data (in many cases, velocity data stopped farther up-valley than the terminus location- e.g. the blue circles from my first figure would be sitting over nothing some years). Buffer zones under these conditions extracted values of Nil, and left the velocity data relatively sparsely populated. While most of the data was well populated for retreat values, it was somewhat difficult to find glaciers with a full 8 years of velocity data. As a result, all of these correlations are performed on velocity data that is more sparsely populated than I would like. Several time series for the glaciers are missing many years of velocity data that could be holding important information.
I would say that overall I am disappointed with how much I got done in exercise 2, though I think in order to do this project correctly it is important to look closely at the data I produce before asking totally new questions. I got stuck in a bit of a rabbit hole looking at all the data I produced and took it out of spatial context. Plus, the project is pretty fluid since it was just generated for this class, so I am free to investigate things that capture my interest rather than a specific hard-nosed question.
Looking forward towards exercise 3, I now want to begin investigating varying buffer sizes, calculating and observing Rate of change (first derivatives) as well as possibly time lags (though looking at the time series data, I doubt time lags are going to have a big effect on any of the correlations). I want to do more spatial cross correlation, in the following exercise to investigate these factors as well (Perhaps IDW, and more Morans I on varying radii).
For Exercise 2, I continued my investigation into the relationship between benthic abundance and total organic carbon (TOC) between Point Conception, CA, and the Canada border with a special emphasis on a region off Newport, OR. For the Newport region, I hypothesize that there is a seasonally fluctuating positive relationship between benthic organism count and TOC. I further hypothesize that commercial fishermen tossing bait overboard, river discharge, summer upwelling, and the dumping of dredge materials outside Yaquina Bay impact this relationship. To investigate my hypothesis, I conducted geographically weighted regressions (GWRs) on benthic box core data from five months (April, June, August, October, and December) in 2011. I have ~12 samples from each of the months. Although the analysis did show some interesting results, I am skeptical of these results due to the lack of variation between data taken from each month and small sample size. For all months, TOC levels were less than 1% and mean organism abundance was about the same. I will run the samples through a t-test to determine if there is any statistical difference between samples from each month. Also, I discovered the GWR requires a large sample size (in the hundreds at least) to produce trustworthy results. Accordingly, I expanded my sampling to a larger area off Newport and included monthly data from all sampled years (2010-2016) so that I could have 100+ data points for each month (April, June, August, October – no December data was available for other years). The results of my analysis on the expanded area differ from my results from the limited region, suggesting a negative relationship between organism abundance and TOC for some months.
These conflicting results led me to expand my investigation area even further; I looked at the relationship between organism abundance and TOC for the entire Oregon Coast and the entire sampling area, with mixed results. Because of my inconclusive results, I believe that I need to take a more focused approach with regard to the species I’m investigating. My advisor, Dr. Sarah Henkel, suggested that I look at the relationship between the ratio of polychaetas (bristle worms) to crustaceans and TOC in order to investigate potential polluted areas. Previous researchers have determined that the ratio between polychaetas and crustaceans can be used as an indicator of pollution. This line of inquiry could help me to identify areas where increased TOC results from pollution (like the dumping of dredge materials and agricultural discharge). I also decided that I need to spend some time reading more papers about benthic TOC levels in coastal areas in order to understand why TOC on the Oregon coast is so much lower than in other areas. One hypothesis is that benthic TOC is low because the waters off the coast of Oregon are incredibly productive – especially during summer upwelling. Huge plankton blooms occur, and those plankton are preyed upon by other organisms before they die and sink to the bottom. Oregon’s rivers also transport large quantities of nitrogen and carbon into coastal areas (especially during the fall-winter rainy season). However, the same storms that cause nutrient transport also cause high wave energy and mixing along the coastline. Transported nutrients therefore remain suspended rather than settling in the benthos. Previous researchers have described both a shift in benthic organismal community structure and a “TOC horizon” off the Oregon coast at 90m depth. Perhaps the coastal processes of upwelling and wave action (among others, i.e. longshore transport) are less impactful on benthic TOC accumulation at that depth.
For this blog post, I will explain the work I’ve done so far using figures and then muse over the process of conducting geographically weighted regressions, keeping my hypotheses in mind.
For my initial analysis, I examined 12 sample sites off Newport, OR, in five months (April, June, August, October, and December). All of the samples were in slightly different locations because of course it’s impossible to get a boat back to the exact same spot. The points depicting the sample sites above are placed at the centroids of the true same points. Not all of the stations had samples for each of the five months, but most of them did. I ran a geographically weighted regression analysis for each month using a fixed kernel size based on AICc (Akaike Information Criterion). I would have preferred to run analysis with a fixed bandwidth parameter using the 30 nearest sample points to a given sample, but I could not use that approach because I had an n value of ~12/month.
In order to understand my result, I graphed the GWR regression coefficients and r2 values for each site (note than MB40 only has 4 months of data and NS50 has three months of data). The graphs reveal a spatially complex and seasonal relationship between organism count and TOC within the sample area. Overall, I saw the strongest positive relationships between organism count and TOC in April and August. The largest r2 values were associated with April (~.6 at some sites) and June (~.7 at some sites) sampling. These results could support my hypothesis that discarded bait from the commercial Dungeness crab fishery in the winter and early spring causes and increase in TOC and organism abundance. Also, the results for August suggest that organism counts increase as a result of spring and summer upwelling (increasing organism abundance). However, as I mentioned earlier, I am skeptical of these results because of the small sample size and relatively small differences between TOC and organism count between months. For instance, I believe the August results might be skewed by one sample site (BB50) that has over 200 bristle worms (Spiophanes norrisi) in a single box core.
To increase my sample size, I included samples taken from 2010-2016 off Newport, OR. Unfortunately, the only December samples that I have are from 2011, and so n=12 even in my extended analysis. The other sample months were April, June, August, and October. For this analysis, I used an adaptive GWR based on 30 nearest neighbors to each sample point. While my analysis of the smaller area only returned one negative coefficient (August 2011 @ NS50), the results from each month in my larger analysis area returned both positive and negative coefficients for each month. I created a box and whisker plot of the results to visualize this relationship more clearly.
I also created box and whisker plots of the organism count and TOC data from each month from the extended Newport sample area (pictured above).
My results show seasonal complexity, with peak mean GWR regression coefficient, mean organism count, and mean TOC occurring in August. Notably, all months included both positive and negative GWR coefficients. Both April and October have negative coefficient outliers. Also, as you can see the TOC box and whisker plots are somewhat ugly as far as having non-normal distribution. A few “next steps” will be to explore methods of data transformation and investigate non-linear relationships or relational thresholds within the organism count/TOC correlation. Dr. Henkel told me that she usually uses a fourth root transformation or organism count data. I suspect that if there is a linear relationship between organism count and TOC, that relationship might only hold true within a range of values. Specifically, I hypothesize that extremely high TOC might be correlated with decreased organism count above a certain level.
To provide context for my Newport area results, I conducted an adaptive GWR using 30 nearest neighbors on an EPA 2003 dataset that ranges from Point Conception to Canada. Unfortunately, I do not know specifically what months the data was collected, but I know that it was taken some time in the summer or early fall. I will dig into the metadata to see if I can find that information. In any case, the results allowed me to see the complexity in the relationship between organism abundance and TOC along the coast.
Red values represent positive GWR coefficients and blue values represent negative GWR coefficients. Because I know that the relationship between organism count and TOC varies seasonally, some of the apparent spatial variability depicted above could actually be seasonal variable – more investigation is required.
My GWR Process
For me, the process of conducting GWR in ArcMap highlighted the impacts of sample size, spatial scale, and adaptive vs. fixed kernel size on results. When I looked at a small sample size (off Newport), the resulting GWR coefficients were largely positive. However, when I expanded the sample area, I saw both positive and negative coefficients within each month. I also noticed that the settings I selected within the GWR tool in ArcMap caused variation in the results. This observation provided me with an example of why it is so important to select the appropriate statistical methodology based on the type, spatial scale, distribution, and sample size of your dataset. For my analysis, I found the adaptive 30n kernel to be most appropriate and useful.
As a reminder, my ultimate research goal is to quantitatively characterize salt marsh volumetric change from 1939 to the present within Oregon estuaries using a combination of sediment cores, for which I have excess 210Pb-derived sediment accumulation rates, and historical aerial photographs (HAPs), from which I plan to measure horizontal expansion and retreat. For my Exercise 1, I successfully georeferenced and digitized HAPs from Alsea Bay that are roughly decadal, spanning 1939 to 1991. The next major step is analyzing my data using the USGS Digital Shoreline Analysis System (DSAS; Himmelstoss et al. 2018). However, prior to analyzing my data in the framework, I must first tie up a number of loose ends; I will explore the following questions for my Exercise 2:
What is an appropriate method of quantifying uncertainty associated with each salt marsh outline?
How should shoreline be defined? More specifically, to what degree should I incorporate channels into my digitization of each layer?
How should I define a modern shoreline?
Step 1: Determine Uncertainties
Each digitized layer must have an associated uncertainty to be run through DSAS. Moore (2000) details all possible errors associated with analyzing aerial photography.
Radial and tangential distortions from camera lenses and film deformations caused during collection, processing, and storage combine to produce significant image space distortions. Distortions are theoretically accounted for by transformations applied to georeferenced images. I used a second-order polynomial transformation which is appropriate for HAPs. However, errors exist associated with each ground control point and without ground-truthing, which is beyond the scope of this study, it is practically impossible to determine the accuracy of my georeferencing. As others have done, I will report the root mean square error (RMSE) associated with each georeferenced aerial image.
Moore (2000) additionally details that objects may appear displaced from true ground positions due to differences in relief, distance from the center of the photo, camera tilt, flight altitude, and atmospheric refraction. In my HAPs, errors associated with object space displacements are particularly great along the upslope edges. However, as I wrote in my Exercise 1 Blog Post, I am not interested in retreat of the upslope edge. Laura Brophy and her colleagues have been working on this issue (Brophy et al. 2019), and I do not believe that the rate of retreat will be significate in my timeframe of interest. Currently I have digitized the upslope edge of all fringing marsh using PMEP’s (Pacific Marine and Estuarine Fish Habitat Partnership) Current Wetland Extent Map. In the future I might either use polylines rather than polygons, or I will mask out the tree line in the finished product.
Other errors are associated with interpretation and digitization of the features of interest (Moore 2000). Others have simply reported the scale at which their images have been digitized. I considered comparing a modern digitization completed by myself to PMEP’s Current Wetland Extent Map; however, this map is primarily based on LiDAR imagery with 5 m resolution, which is much coarser than the resolution of my imagery. Thus, to account for errors associated with digitization I will adopt the common method of assuming 1 m of uncertainty based on previous studies (Ruggiero et al. 2013).
Based on Ruggiero et al. (2013) I plan to characterize error for each year using the equation:
where Up is the total uncertainty of the shoreline position, Ug is the uncertainty associated with georeferencing (RMSE of the control points used to georeference each image), and Ud is the uncertainty of the digitization (1 m). Uncertainty associated with NOAA T- and H-sheets is considered to be 10 m (Ruggiero et al. 2013).
Step 2: Reduce Channel Complexity
Though I have not assessed the impact of scale on changing the overall area of each marsh platform, I hypothesize that smaller spatial scales would result in decreased volume as more and more tidal creeks are incorporated. Because I have digitized my salt marshes at the smallest spatial scale possible for each photo and the highest quality HAPs are 1939 and 1952, these may erroneously appear relatively too small due to higher channel density than the lowest quality HAPs, 1983 and 1991. Thus, I must decide how to correct for differences in salt marsh channel digitizing due to differences in spatial and spectral resolution between photographs as this alters the overall area of the salt marsh complex.
Preliminary results suggest that resampling the cell size of the highest resolution photograph (1952, 0.2 m) at the lowest resolution photograph (1983, 0.7 m) resolution using the ArcGIS Pro Raster Cell Resample tool (Nearest Neighbor and Cubic resample type) does not produce a noticeable difference. Thus, it appears as though differences in spectral resolution primarily impact apparent differences in image qualities, and thus ability to distinguish tidal creeks. Unfortunately, because no meta data besides date was provided by the UO Historical Photography Library, assessing spectral quality is outside the scope of this study.
My next attempt to standardize channel digitization was to limit channels to widths greater than or equal to the uncertainty associated with each image. This was performed using the ArcGIS Pro Buffer tool. I simply selected the marsh polygon of interest, selected the Buffer tool, specified the buffer distance as the error (e.g., 1 m), then selected that buffered polygon and buffered it again, specifying the buffer distance as the negative error (e.g., -1 m). Comparison of the second buffered polygon indicates that little to no resolution was lost on the marsh edges but that the channels were appropriately limited. I then deleted the original polygon and the first buffered layer.
Upon completion of this first round of Buffering, I realized that because the error is different between each year, the channel density is still different. Thus, I repeated the first step using the worst uncertainty estimate of 3.8 m. At this point all marsh outlines should have a channel density not primarily driven by differences in spectral resolution.
Step 3: Determine a Modern Shoreline
As stated previously, the PMEP map uses LiDAR with a resolution of 5 m, which is too low for comparison with my HAPs. An OSU MS thesis from 2004 (Russel Scranton, 2004) digitized Alsea Bay using digital ortho quads from 2001 and ground-truthing with RTK GPS. I incorporated this layer, though it required some edits made using the original 2001 digital ortho quads. Because I am comparing horizontal growth rates with sediment core derived vertical accretion rates (and my cores were collected between 2013 and 2017) I then needed to decide what imagery to use from the last ~20 years. I settled on 2009 and 2018 as these years have aerial imagery taken at relatively high resolution (< 0.5 m). I digitized these in a similar manner to how I digitized my HAPs.
Step 4: Merge Data into a Single Feature Class
DSAS requires data be merged into a single feature class using the Merge tool in ArcGIS Pro. This file must also have two columns in the attribute table noting the date and uncertainty of each shoreline. I specified Jan. 1 for each date as I mostly only had the year of collection, and I specified the uncertainty using the above method. Originally, I merged my polygon data but after realizing DSAS only runs with polyline data, I converted each individual layer using the Polygon to Line tool in ArcGIS Pro, and then re-Merged and edited the attribute table of my shoreline layer. Before placing the data into a personal geodatabase that will be accessed by DSAS in ArcMap, I made certain my layer was in meters.
Step 5: Initial DSAS Run
DSAS requires a baseline layer for orientation when casting transects. I decided to use my 1939 polyline buffered at 10 m as the baseline. In ArcMap, I specified my baseline, making certain that the program would cast offshore. I then performed an initial casting of transects using a 10 m search radius, 10 m transect density, and a value of 50 for transect smoothing. After calculating transect statistics, I colored my transects based on rates of growth or erosion.
Final results related to this project are related to estimates of uncertainties and the spatial resolution for each year. Uncertainty ranged from 1.0 to 3.8 m amongst shorelines digitized from the last century. These values are actually quite low compares to other similar studies (e.g. Schieder et al. 2017). Because Oregon salt marshes are typically much smaller than those on the US East and Gulf and European coasts, my georeferencing RMSEs are likely smaller as images need not be stretched over greater distances.
The 1887 shoreline has an uncertainty of 10 m but I have not year included this in the analysis. Because the change in shoreline shape between 1887 and 1939 is so stark, it should be analyzed separately from the 1939 to present data.
The spatial resolution ranged from 0.30 to 0.73 m. Estimating spectral resolution is beyond the scope of this study.
Table 1. Uncertainties (calculated based on georeferencing RMSEs and digitization uncertainty of 1 m) and spatial resolution for each year.
Total Uncertainty (m)
Cell Size X (m)
Cell Size Y (m)
The initial DSAS run clearly highlights that though there are a few areas of accumulation, there exists primarily areas of erosion in the Alsea Bay salt marshes. It appears as though the northern edge of the largest salt marsh island has been experiencing the greatest rates of erosion. This pattern is likely related to the diking of the northern portion of fork of the river in the middle of the century. This would have reduced connectivity and thus sediment transport to this portion of the bay, resulting in erosion under rising relative sea level. The northern portion of the fringing marsh appears to have experienced the slowest rates of erosion. The reason for this is unclear but potentially related to oscillating periods of accumulation than erosion as evidenced by a detailed look at the different shoreline layers.
Despite these initial exciting results, there are clearly many areas where this initial DSAS run can be improved and these are both highlighted in the image and briefly outlined below.
As a general note, I am very happy I chose to move forward with just Alsea Bay salt marshes for this project, rather than georeferencing and digitizing all of my salt marshes at once. Determining all of the little quirks of the Editing tools in ArcGIS Pro and the specifications of DSAS on one dataset will allow me to move much faster through successive datasets. For instance, I did not realize that DSAS required polylines rather than polygons. Going forward I will digitize all of my marshes using polylines rather than polygons and I believe this may actually save me time since I will not have to digitize any upslope edges.
The methods I used for determining uncertainty seem to be universally accepted given the difficulty of estimating true accuracies of shoreline positions back through time. I will continue to use this method for the remainder of this project.
The Buffering tool saved me a lot of time. Considering a digitization for a single year takes me approximately one day to complete, not having to re-digitize at a universal resolution or to a specific channel width allowed me to make more progress in other aspects of Exercise 2.
I am really excited to move forward with the DSAS program. For my next Exercise, I plan to play around with the DSAS settings to improve the transect projections. Depending on the difficulty of this, I may also have time to being incorporating my sediment core data. Below are some initial areas for which the DSAS projection could be improved.
Remove or mask out all upslope edges since these are not of interest and all digitized using the same PMEP boundaries.
Remove or mask out all complexity within the salt marsh complexes, especially within the northern portion of the larger island and the southern portion of the fringing marsh.
Play around with the transect search distance to remove transects bridging the gap between the largest and smallest islands. This may require running each marsh complex separately (which I would rather not due so that all transect symbology are on the same scale).
Play around with transect smoothing to reduce the number of intersecting transects in areas with complex shoreline.
Brophy, L. S., Greene, C. M., Hare, V. C., Holycross, B., Lanier, A., Heady, W. N., … & Dana, R. (2019). Insights into estuary habitat loss in the western United States using a new method for mapping maximum extent of tidal wetlands. PloS one, 14(8).
Himmelstoss, E.A., Henderson, R.E., Kratzmann, M.G., & Farris, A.S. (2018). Digital Shoreline Analysis System (DSAS) version 5.0 user guide: U.S. Geological Survey Open-File Report 2018–1179, 110 p., https://doi.org/10.3133/ ofr20181179.
Moore, L. J. (2000). Shoreline mapping techniques. Journal of coastal research, 111-124.
Ruggiero, P., Kratzmann, M. G., Himmelstoss, E. A., Reid, D., Allan, J., & Kaminsky, G. (2013). National assessment of shoreline change: historical shoreline change along the Pacific Northwest coast. US Geological Survey.
Schieder, N. W., Walters, D. C., & Kirwan, M. L. (2018). Massive upland to wetland conversion compensated for historical marsh loss in Chesapeake Bay, USA. Estuaries and Coasts, 41(4), 940-951.
Scranton, R. W. (2004). The application of Geographic Information Systems for delineation and classification of tidal wetlands for resource management of Oregon’s coastal watersheds.