GEOG 566

         Advanced spatial statistics and GIScience

April 7, 2017

Addressing Spatial Patterns in Multispectral Imagery of SW White Pine Seedlings Grown in Common Garden Boxes

Filed under: 2017,Final Project,My Spatial Problem 2017 @ 8:49 am

1. The research question I explored in GEOG 566 is: “What relationship do spectral reflectance signatures of southwestern white pine (P. strobiformis) seedlings in common garden boxes have with box number and distance from center of box?”

Another way to word the question is : “How do spectral responses differ with the boxes that seedlings are grown in or to the position of the seedlings within the box?”

2. My raw data consist of 500 photos taken from a UAV of common garden boxes in northeastern Arizona. I used Agisoft Photoscan to compile the images into a 5 layer orthomosaic (Figure 1) where each layer represents a discrete band from the multispectral sensor (Table 1).  Processed mosaic images are georeferenced using RTK GPS coordinates for targets which were arranged around perimeter of the AOI. Once georeferencing is completed, the stacked orthomosaic is exported from Photoscan as a TIFF file which can be viewed and manipulated further in ArcMap or R. Two additional spectral index layers (NDVI and TGI) were easily created from the other bands using raster algebra (Figure 2).

Analyses were conducted on the orthomosaic raster which is in .TIF format and is 336MB in size. The extent of the scene is about 10m x 29m and includes 1697 seedlings. The pixel width of the image (spatial resolution) is about 5mm. This fine scale allows for individual leaves to be detected in the image.

Figure 1: True color composite orthomosaic image of common garden boxes containing southwestern white pine (P. strobiformis) seedlings in Kaibab National Forest, Arizona, USA.

Table 1: Micasense Rededge multispectral sensor band designations and spectral information.

Figure 2: Normalized Differential Vegetation Index (NDVI) layer for one common garden box

3. Based on visual analysis of NDVI hot spot analysis (Figure 3), there appears to be some patterns both at the box and plot levels. Specifically, the healthiest seedlings are mostly grouped in the center of boxes and the least healthy ones are mostly grouped around the edges. Also, the healthiest and least healthy seedlings seem to occur in different boxes nearly all of the time.

Based on this, my hypothesis was that spatial statistics would reveal patterns between spectral reflective signatures of individual plants and individual boxes and/or the location of the plant within its box.

Furthermore, I hypothesized that by regressing on these spatial variables, the distribution of spectral responses would resemble more what we expect in a common garden experiment: randomness.

Figure 3: Hot spot analysis reveals visual patterns in  mean crown NDVI both across and within boxes.

4. Following the inital hot spot analysis of mean crown NDVI in ArcMap, I used ordinary least squares (OLSR)  to view the distributions of my spatial variables compared to NDVI. The results support the notion that hypothesis 1 is correct. Next, I used geographically weighted regression (GWR)  to test the first hypothesis and weigh the variables. Because the multiple regression seems to bring out the randomness in the spectral data, hypothesis 2 was also well supported.

5.  The results of my OLSR analysis (Figure 4) supported my first hypothesis. The box number (BOX_) variable has an obviously non-normal distribution, and the distance from center of box variable (BOXCENTER) has a skewed distribution. The other two variables, distance to nearest neighbor (NN) and distance to plot center (PLOTCENTER) appeared to be normally distributed. As a result, I was able to confidently move forward and look into the second hypothesis only for the two spatial variables mentioned (BOX_ and BOXCENTER).

Figure 4: Distributions of 4 spatial variables vs. NDVI.

To investigate my second hypothesis I created GWR maps of my seedlings, first with BOX_ as a lone explanitory variable (Figure 5), then with BOX_ and BOXCENTER together (Figure 6). The resulting data are much less obviously spatially autocorrelated, suggesting that the second hypothesis could also be confirmed.

Figure 5: GWR of NDVI using box number as an explanitory variable reveals a large reduction in the number of clumped outlier datapoints and an increase in the apparent ‘randomness’ of spectral signatures

Figure 6: Though the effect is less pronounced than the box number regression, adding distance to box center to the GWR did seem to further increase randomness, especially at the east and west extremes of the area of interest.

6. My analyses are significant for a few main reasons. For my own research, they mean that I now have a protocol for processing this sort of dataset in a way that allows me to account for spatial patterns. I did not expect to see a box effect of this magnitude but I am much more equipped to address it moving forward.

For the larger project I am part of, it means that we need to account for (or at least test for) the box effect in the context of all analyses. If the boxes are really as different from one another as my spectral data suggest, then some of the assumptions other members of my research team are making could be invalid. One way to investigate this could be testing soil moisture content across many boxes to look for differences.

For science as a whole, there is promise that I will be able to reliably phenotype seedlings based on relative drought resistance in the next two years. Even more exciting, my workflow can also be used in other experiments with seedlings grown in common garden boxes, including disease resistance screening. If effective, these techniques could greatly reduce the man hours required to conduct these screenings and allow for larger, more comprehensive experiments without requiring as much funding.

7. I learned immensely in this course, especially about tools in the “spatial statistics’ toolbox in ArcMap. I had not conducted hot spot analysis, OLSR, or GWR with my own data before. After doing so, I feel much more capable of carrying out and interpreting these.

Also, I was able to create a complex and versatile R code over the course of the term. Within R I learned to test for and execute principal components analysis (PCA) and to automate many of the step in my processing workflow. Without a doubt, the latter will save me hundreds of hours of time. For that, I am simultaneously, impressed, proud, and grateful.

8. Though I knew about the spatial statistics toolbox, I did not know how any of its functions worked nor did I know how to interpret the results. After completing this term, I know how the data need to be formatted to analyze patterns, map clusters, and model spatial relationships in ArcMap. Also, though I still have much to learn, I can now explain what the results of some of these analyses mean within the context of actual data.

In regards to learning about how to use R, I have been in an exponential period of learning since I started using it this past February. Just this term I have been introduced to several packages that have exciting applications within my research interests. For more details about how I used R to automate my data analysis, see my tutorial 2 here.

9. The comments I received about my tutorials and presentations have been helpful for me to stay on track with my analysis. Periodically, Dr. Jones re-oriented me on the specific question I was asking and how I could refrain from spurious analyses. Because I have so many variables and no experience doing this type of work, that orientation was especially useful and in the end instrumental in the volume of what I was able to accomplish.

My peers were helpful because they gave me a sounding board to bounce my ideas off of in the small group presentations. By reading over the synopses of my short talks I was able to get an idea of how well I explained my study and results. I’ve tailored this final edit to my spatial problem with their comments in mind in order to give the best final presentation possible.

Print Friendly, PDF & Email

No Comments

No comments yet.

RSS feed for comments on this post.

Sorry, the comment form is closed at this time.

© 2019 GEOG 566   Powered by WordPress MU    Hosted by