My goal for this class was to learn what tools available in ArcGIS’s Spatial Statistics toolbox might be useful in helping to answer my research questions regarding spatial-temporal relationships between mountain pine beetle outbreak\spread and coinciding environmental characteristics such as topography, climate, and forest attributes. I quickly learned that the spatial statistics tools in ArcMap are not suitable for my data. All tools expect vector data as inputs and I am working with raster image data. I can do conversions from raster to vector, but it really doesn’t make sense computationally and theoretically. Also, I found that my image data needed a lot more pre-processing to get it to a point where it could be analyzed than I originally thought. I spent the majority of the class time preparing my data, which included: georegistration, disturbance detection, and disturbance causal agent identification. Once I had completed these steps for a pilot image scene I was able to do a quick look at outbreak elevation histograms through time, and also annual climate prior to and during the outbreak.
The remainder of this post will walk through the steps I took to process the imagery and present some initial findings on mountain pine beetle outbreak and environment relationships.
The first step in image preprocessing was to properly georegister mis-registered images. Please click on the images below to see an example of a before and after image registration correction.
Figure 1. Satellite image before and after georegistration correction (click images to see animation)
The registration process was coded and automated using the R “Raster” package. The first step was to sample the image with points and extract small image chips or windows from around the point in the reference image and the misregistered image (see figure below).
Figure 2. Sample reference and misregistered image and extract image chips from around the sample points.
Within the registration program, for each sample point, the misregistered image chip is “slid” over the reference image chip at all defined x and y offset distances and cross correlation coeffiecients are calculated for each shift based on the intersecting pixel values. Please click on the figure below to see an example of the moving window cross correlation.
Click image to see animation
Figure 3. Animation example of moving window cross correlation analysis between reference and misregistered image chips for iterative xy offset shifts.
A cross correlation surface is produced for each sample point. The apex of the conical peaks represent the offset that is needed to correct the misregisted image. A 2nd order polynomial equation is created from all legitimate peaks and used to warp the misregistered image into its correct geographic position.
Figure 4. Example cross correlation surfaces showing the best offset matches between the reference image and misregistered image chips
With all images in their correct geographic position, the “LandTrendr” change detection program was applied to the long time series of Landsat imagery to identify pixels that have been disturbed. A discriminant analysis of empirical variables related to spectral, shape, and topographic characteristics of identified disturbances was conducted to predict disturbance agent from a training set of labeled disturbances. The figure below depicts a mountain pine beetle outbreak identified in central Colorado. Click on the image to see the outbreak start and progression (colors represent magnitude change to forest: low-high\blue-red)
click image to see animation
Figure 5: Mountain pine beetle outbreak spread as captured by annual Landsat satellite imagery
From the outbreak progression video above, you can see that the outbreak appears to move up slope. To find out if this is truly the case I extracted elevation values for all insect affected pixels per year and plotted the histogram of both elevation value density and frequency. Please click on the images below to see the shift in elevation and area affected as the outbreak progresses.
Figure 6. Animated histograms depicting the progression of mountain pine beetle progression up slope.
I was also curious about what the annual climate was doing before and during the outbreak. I extracted PRISM climate data from 1985 to 2008 in the region of the outbreak and plotted it with the count of insect-disturbed pixels. The figure shows that maximum and minimum annual temperature begin to increase 1 to 2 standard deviations from mean about 5 years before the outbreak really takes off. This corresponds with a 3-4 year drop in annual precipitation. These conditions could have drought stressed the forests and provided a highly productive climate for the beetle to reproduce multiple times in a season and avoid freeze-kill.
Figure 7. Graph showing annual deviation from mean for PRISM climate variables and insect-disturbed pixels for 23 years.
In closing, I found the ArcMap spatial statistics unable to work with my raster format data, but was able to make a lot of progress in data preparation and analysis and exploration of satellite image detected insect outbreaks and corresponding environmental factors.
Hi Justin,
Wonderful animations! I’m wondering what the historic outbreak ecology record is like and if there is any literature linking decadal (or shorter/longer) oscillations to outbreak?
Kate
Great work Justin–very cool animations.
I wonder if growing degree days (or days above a known beetle kill temperature) may help to include data that relates more closely to the beetle’s life history. It looks like high T and low ppt has an effect (forest stress), but it seems to fully capture the infection process you’d want to include the beetle as well. Granted, it’s hard (impossible?) to get that type of temporal resolution from PRISM data…could you correlate weather station data to monthly PRISM averages? Depending on the density of weather stations, you might be able to do some interpolations of GDD with elevation. Thoughts from a former beetle biologist…
Hi Justin,
First off, congratulations on overcoming your data preparation problems. I can certainly sympathize. I’m very impressed by your solution (and that must certainly have taken some awesome coding skills!). I would have expected a high degree of error as adjacent pixels are probably fairly similar (Tolber’s Law?) to each other, but your fit curves look very tight and the final product looks quite accurate, so well done. Were there some error/uncertainty problems or concerns?
Regarding your actual spatial research question, I was thinking that your question might be able to be refined. In looking at that elevation pattern you picked up, it seems from the image like the only available, originally disease free area for the beetles to spread into was higher up the mountain. If they were going anyway, it had to be to higher elevation.
So, for your question, perhaps you could first define the potential areas where disease could spread into? Then asks questions on whether there’s a preference within that potential habitat or if rate of spread is slowed by certain characteristics? The main point is I feel it’s important point to first define what potential new expansion areas were available.
That said, you have a very striking visual and I’d be excited to see such maps for other areas and then perform analyses of how different spreads are different. Very interesting first step!
Justin,
Very nice results. Your histograms imply to me that the outbreak began as two separate patches at different elevations, which coalesced as the outbreak expanded. However, this is not evident in the animation of the landscape. It makes me wonder whether you could use this example, and others, to develop an algorithm for identifying the characteristic spatio-temporal pattern of an insect outbreak. Does it look different from other spatio-temporal processes, such as a wildfire (much faster), or urban expansion (slower)? Adjusting for the differences in timing, do insect outbreaks have distinctive spatio-temporal patterns compared to wildfire or urban expansion?
Julia