Research Questions

Exercise 1: What are the spatial patterns of vegetation species distribution across sampled points?

Exercise 2: Is there spatial cross-correlation between pairs of plant species presence along a sampled transect? Is there spatial cross-correlation between the presence of a given plant species and elevation along sampled transects?

Exercise 3: How did vegetation presence change between 2015 and 2019? Where was there a gain or loss of vegetation, and what areas remained vegetated or unvegetated?

Datasets

For the first component of Exercise 1, I examined a 2021 elevation and vegetation dataset. Elevation was collected with an RTK-GPS and vegetation species presence and maximum height were recorded at each point. Sampling was done every 50m on a grid. For the second component (autocorrelation) of Exercise 1, and for Exercise 2, I used an elevation and vegetation dataset from 2021 resampling of permanent transects (which have been sampled nearly annually since 2009). The same sampling method was used as the previously described dataset, except sampling was done every 1m along each 50m transect. Additional data was collected but was not used for these analyses.

For Exercise 3, I used 4-band multispectral aerial imagery from with 0.25m resolution from 2015 and 2019.

Hypotheses

I predicted that broadly, the spatial pattern of plant species would be clustered at the site scale, due to differences in abiotic conditions (i.e. salinity and elevation). I predicted that there would be some differences between spatial patterns, which I expected is due to smaller scale differences in abiotic conditions as well as biological interactions (not being investigated). In terms of site-wide vegetation change, I predicted there would be a net gain in the extent of emergent high marsh over time, if with the restoration of tidal influence there has been sufficient sediment availability for vertical accretion to occur via a positive feedback loop between accretion and vegetation growth (Kirwan et al. 2013). (I did not end up looking at specific habitat types or plant communities, but still would predict vegetation gain on the mudflat in higher elevation areas.)

Approaches

Exercise 1: point pattern analysis (average nearest neighbor) and autocorrelation

Exercise 2: cross-correlation

Exercise 3: confusion matrix and change detection map

Results

In Exercise 1, I produced maps displaying the presence and absence of two marsh plant species, marsh jaumea and saltgrass. I produced statistical relationships for the average nearest neighbor analysis. For this analysis, I selected a subset of points so that there would not be gaps in the sampling coverage; the sites sampled were not contiguous, creating gaps that would have affected results. The observed mean distance for saltgrass is 48.61m, which reflects the 50m sampling grid. The observed mean distance for marsh jaumea is 70.14m.

For the autocorrelation component, I produced plots. For one of the transects, I found significant autocorrelation for lags 1-4. I then appended two transects from a different restoration unit to increase the sample size to 100 points, and found significant autocorrelation at all lags, decreasing over space.

For Exercise 2, I produced cross-correlation plots for the relationship between saltgrass and jaumea presence/absence along two transects and between saltgrass presence/absence and elevation. For the pairs of species, there was no significant cross-correlation on one of the transects. For the other transect, there is some significant cross-correlation (maximum ~0.35) for lags -4 to -13, decreasing over space. I believe this indicates significant cross-correlation of saltgrass to the left of a point with jaumea. For elevation and saltgrass, there is significant cross-correlation between lags -2 to 2, and the plot is fairly symmetric. For the other transect, there is significant cross-correlation from lags -4 to 6. These findings make sense to me, as I tended to see saltgrass at higher elevations within the tidal frame.

For Exercise 3, I produced a map of vegetation change, and a sort of confusion matrix (summary statistics for the percent represented by each category). I believe that there are inaccuracies in this analysis (see next section), but it does appear that by 2019 there was some vegetation colonization surrounding vegetation that existed in 2015.

Analysis Learnings

In Exercise 1, the average nearest neighbor analysis didn’t turn out to work well for my data, due to gaps between sampled areas, as well as observer-determined sampling points (50m grid). For the autocorrelation analysis I used the transect data because pulling out a transect from the grid wasn’t enough data to use. I learned that the presence (or absence) of saltgrass at one point did tell me something about the presence (or absence) at the next few points for one transect, and when I appended transects for another unit, there was autocorrelation at every point.

In Exercise 2, I used a very limited set of data (one pair of species for two transects, and one species and elevation for two transects). The plots look relatively different for the sets of transects, which I expected as I consciously chose two that would be different (different habitat, one with a transition from mudflat to vegetation, etc.). There was significant cross-correlation between saltgrass and elevation, and I’m interested in investigating relationships for more species and transects. Additionally, I think using salinity instead of elevation will be informative.

Exercise 3 was primarily useful for learning the process of a basic change detection analysis and becoming aware of data issues I’ll need to address. For example, some parallel patterns of vegetation gain and loss on the northeastern island indicated that the channels are not lining up well and I will likely need to do new georeferencing. Additionally, I believe there are areas of the large mudflat that were categorized as vegetated but are actually algae, showing up as a loss of vegetation.

Significance

The results from autocorrelation and cross-correlation in Exercises 1 and 2 show some promise for predicting the presence of vegetation. If this is the case upon further investigation, this may be useful in modeling future post-restoration trajectories (i.e. under different sediment accretion scenarios, where would vegetation be predicted to colonize?). Additionally, cross-correlation between species may provide information on common plant associations. I think that these analyses are likely most useful as initial steps that might inform future analyses.

The “confusion matrix” and map of change detection give a site-wide view of vegetation change (though currently without the nuance of habitat type). Once some issues are addressed and new results are produced, the map will be a helpful visual of patterns of change over time. For example, a freshwater wetland transitioned to mudflat post-levee breach in 2009, and there need to be elevation gain for vegetation colonization to be possible. Maps such as this produce a visual of whether this has occurred, whether it’s occurring in specific areas and/or as a result of known or unknown processes (i.e. the eastern side where it was predicted there would be more sediment input). This analysis contributes to monitoring efforts, and can inform resource managers with adaptive management decision-making (I.e. could plantings or sediment application be warranted)? This research fits into a larger monitoring framework at the site. Additionally, monitoring contributes to knowledge about the time frame and trajectory of restoration, which may inform the design of future restoration projects.

Software learning

I used ArcGIS Pro for data visualization, point pattern analysis, and the confusion matrix and change map. The steps involved in these analyses introduced me to more geoprocessing/spatial analysis tools in Arc. I used R for data manipulation (into the presence/absence and elevation format I needed) and for spatial autocorrelation and cross-correlation, which were new R functions for me. I did not end up using Python or Modelbuilder in Arc.

Statistics learning

I learned that point pattern analysis is a good option for presence absence data, though my sampling points being observer-determined hindered the average nearest neighbor analysis (as well as gaps between sampling sites). I chose spatial autocorrelation because I have evenly spaced count data, and determined that presence/absence data would indeed work as a count. In the limited bit I’ve heard about autocorrelation in the past, it’s been in terms of checking for violations of model assumptions (I.e. regression analyses), so it was helpful to learn about new applications for univariate analysis.

Cross-correlation was a new statistical method to me. One limitation I found was that I was unable to run this analysis in the instance where a species was present at every sampling point. In working on change detection, I was able to think through ways to deal with the issue of NDVI not being standardized between years, and using unsupervised classification. I hadn’t run these analyses before, so I learned a lot about the process!

Evolving question

My objective in “My Spatial Problem” was to explore the spatial patterns of vegetation species and communities, how vegetation community structure has changed since restoration, and how this related to geomorphological change via changes in sediment delivery and inundation regimes.

Wow, that was a broad question! For the first two exercises, I ended up focusing on individual vegetation species, rather than tackling any sort of community analysis (to come in future research questions). My original question was missing relating vegetation species presence to a variable B (elevation for exercise 2); I had skipped ahead to broadly stating that I wanted to relate vegetative change to geomorphic change.

My restated questions are: How is species distribution related in space to physical environmental variables (i.e. elevation, salinity, proximity to channels)? How are patterns of vegetation change related to geomorphic change? 

Future techniques

I would like to continue working on change detection, looking more into habitat classification methods. For example, I would like to learn more about unclassified habitat supervision and whether manual adjustments need to be made for areas of potential misclassification (such as algae being classified as veg). In the future, I will incorporate empirical data to classify points on spectral signatures, and then do image classification. Additionally, I’d like to look into other values for classification such as the Soil Adjusted Vegetation Index (SAVI) that adjust for soil reflectance, or other classification methods such as object-based classification. Ultimately, I will want to perform change detection by habitat type (mudflat, salt marsh, riparian floodplain, etc.) to better quantify restoration progress.

I would like to explore dissimilarity analyses, such as Bray Curtis, to look at changes over time in resampled vegetation quadrats. I’ll be exploring ordination techniques once I’m further along in thinking about vegetation community analysis. I’ll also be doing a lidar change detection analysis and relating this to vegetation change (technique options to be explored!).