Sorry for the bad puns in the title… Could not help myself.

QUESTION

How do spatial patterns of black stain root disease infection probabilities relate to spatial patterns of forest management classes? How do these spatial relationships differ between landscapes where similar management classes are clustered and landscapes where management classes are randomly distributed?

TOOLS AND APPROACH

I used neighborhood analysis to analyze the spatial relationship between hot and cold spots of infection probabilities and the three forest management classes simulated in my model (extensively managed plantations, intensively managed plantations, and passively managed old-growth stands).

I used a combination of ArcMap (to perform the majority of the procedure) and R (mostly for spatial data wrangling and analyzing and plotting results).

DESCRIPTION OF STEPS YOU FOLLOWED TO COMPLETE THE ANALYSIS

1. Compare the distribution of infection probabilities between landscapes and management classes

I performed this step in R to determine whether there was evidence of significant differences in the infection probabilities (a) between the clustered and the random landscapes and (b) between management classes both within and among landscapes.

2. Hotspot analysis

Converted my raster data of infection probabilities to point data (in R, using the “raster” package) and perform a hotspot analysis (hotspot tool in ArcMap) (Fig. 1). For the hotspot analysis, I used inverse squared distance weighting to conservatively include trees within hotspots.

I also created polygon shapefiles for areas identified as hotspots to the 99% confidence level in each of the landscapes and calculated the area of these hotspots for each landscape, management class, and landscape X management class.

Figure 1. Hotspot analysis overlaid on forest management classes for the “clustered” landscape. The same analysis was performed on the “random” landscape.

3. Select point for neighbor analysis

For this step, I chose one point in a hot spot and one point in a cold spot for each of the two landscapes to perform the neighborhood analysis. In the future, I would write a script to repeat this procedure with a random sample of a large number of points, but for this exercise, I just used one point for each.

For the hotspots, I only used points identified in hotspots at the 99% confidence level. For the cold spots, I used a point from the 99% confidence level for the clustered landscape, but my only options for the random landscape were points identified at the 90% confidence level. I visually selected the points for analysis (non-random), but I would not do so for my full analysis.

4. Create concentric ring buffers for the neighborhood analysis

I used the Multiple Ring Buffer tool in ArcMap to generate hollow ring buffers at a series of distances around each hotspot and cold spot point selected for analysis (Fig. 2). I then intersected these buffers with the management class shapefile and calculated the proportion of each buffer ring composed of each management class (by area). I plotted these proportions as a function of distance to complete the neighborhood analysis.

Figure 2. Concentric ring buffers used for the neighborhood analysis, with outer ring distances labeled. Example shown for the “random” landscape for both hot and cold spots.

RESULTS

Non-spatial analysis

Before performing the neighborhood analysis, I wanted to know whether there was any difference in the infection probabilities between the two landscapes and the three management classes. Visual assessment of the box plots (Fig. 3) led me to believe that there were significant differences between the means of the infection probabilities between the landscapes and management classes, which was supported by the results of student t-tests (p << 0.01). There were also significant differences in the infection probabilities between management classes both within and among the two landscapes. Since the outputs I am analyzing are “dummy data” because my model is not fully complete, this does not surprise me and I did not perform further, more rigorous statistical analysis. The higher infection probabilities simply reflect the density of trees in each of these management classes. Extensive management had the highest infection probabilities (highest initial planting density, evenly spaced), followed by old-growth (lower density but clustered trees) and finally intensive management.

Figure 3. Comparisons of infection probability distribution between (a) landscapes, (b) management classes, (c) management classes by landscape; (d) the proportion of each management class covered by an infection hot spot in each of the two landscapes (by area).

Neighborhood analysis

For the clustered landscape, the management class in which the point was located made up the largest proportion of the first several distances analyzed (cold spot: old-growth, hot spot: extensive). This makes sense given that the management classes were spatially clustered in this landscape. At a certain threshold, the proportion of this management class started to decrease, as the other two management classes increased in proportion at a similar rate. For the random landscape, the decrease in the starting management class proportion was relatively rapid with distance, and all three management classes converged at their landscape proportion by 150 meters, with some fluctuation (in both landscapes, each of the three management classes makes up about 1/3 of the total landscape).

Figure 4. Neighborhood analysis showing the proportion of each (hollow) ring buffer accounted for by each management class for hot and cold spot points in each of the two landscapes.

Critique of the method – what was useful, what was not?

Most of the drawbacks of this analysis were due to my process and the nature of my data. Because I only used one hotspot and one cold spot point for each of the two landscapes, it is difficult to say much from this analysis. If I were to automate the analysis and run it on a series of random points drawn from the hot and cold spots, I would get a lot more insight as to the patterns of neighborhood effects if any exist. In addition, the “dummy data” used for analysis come from model runs that only have local disease transmission (between neighbors at a radius of several meters). However, the full model will also incorporate long-distance dispersal by insect vectors (on the scale of kilometers), which will likely be much more interesting and less predictable when neighborhood analysis is performed.

Issues with the method itself is that it is quite time intensive – an issue that could be cleared up with a script written in Python or R to automate this process given a set of input files. If there is a way to do this already built into either of these platforms (or Arc/QGIS), I was not able to find it. Also, there is a lack of clear quantitative interpretation for these plots to separate statistically significant variability in the proportion of each management class at each distance from non-significant variability. A means of doing so would enrich the analysis.