For Exercise 2, I wanted to explore how the spatial distribution of land use/land cover (LULC) varied in relation to an artisanal, small-scale gold mine (ASGM). To do this, I took the NDVI change maps which I had generated for Exercise 1 (modified slightly to produce better, more accurate results), as well as my dataset of mapped footprints of known areas of ASGM activity, found the median center of those areas, generated buffers/donuts at 250m apart, from 250m to 2000m, clipped the NDVI change layer to each buffer, and counted the amount of each type of NDVI pixel contained within each buffer. In addition, I performed this same analysis around non-mining villages, to examine how the spatial pattern of NDVI loss changes with distance from the center of non-mining villages. With this data, I could generate a chart showing how the percentage of pixels representing loss or decrease in NDVI changed as you moved further away from the center of mining activity. This examination looked at nine mining villages and nine non-mining villages
In order to perform this analysis, I continued my work using ArcGIS Pro, in addition to Excel for a bit of charting work.
To begin, I imported my NDVI change map, which detailed increases and decreases in NDVI values between imagery taken in April 2018 with Landsat 8, and imagery taken in April 2007 with Landsat 5, representing 11 years of NDVI change. I also imported my shapefile containing polygons which I had digitized over my high-resolution satellite imagery depicting areas of known ASGM activity. With this shapefile, I used ArcGIS’s Median Center tool, which found the median center of the mining areas near the village of Douta (fig. 1). From there, I generated buffers/rings at 250m intervals (e.g. 0-250m, 251-500m, 501-750m, etc.), from 250m to 2000m around this median center. I then used the clip tool to clip the overall NDVI change layer to each buffer, resulting in NDVI change for 0m to 250m from center, NDVI change for 251m to 500m from center, and so on.
Fig 1: Center of the village of Douta with a 2km NDVI buffer; the blue buffer shows the 1750m extent, whose values were subtracted from the 2000m values, to only give values between 1751 and 2000m
Once I had accomplished this, I generated histograms for each buffered NDVI change layer in order to count the amount of pixels contained within each buffer, and assign them to one of two classes: negative values, representing overall NDVI decrease from 2007 to 2018, and positive values, representing overall NDVI increase from 2007 to 2018. I did not account for magnitude of change, as I wanted a general idea of how NDVI was changing from the center. Fig. 2 shows an example of these histograms, specifically for the 500m buffer. The values from the previous buffers, e.g. the one to 250m, were subtracted to only show values from 251 to 500m.
Fig. 2: Douta histogram
I entered all the pixel values for negative change at each distance into Excel, as well as all of the pixel values for positive change, and was able to generate a chart showing how the percentage of the overall pixels in each subsequent buffer from center representing NDVI decrease change over distance (fig. 3)
Fig. 3: Pale blue lines indicate individual mining villages. The dark blue line indicates the average of those villages. Thin red lines indicate non-mining villages. The dark red line indicates the average for those villages.
On the whole, this exercise was useful for illustrating the problem I’m attempting to grapple with. I was frustrated with the lack of Landsat imagery from the late 2000s — I was unable to find any Landsat 5 imagery corrected for surface reflection aside from the year 2007. Additionally, there are problems with comparing this 2007 image to the 2018 image. I found that the rainy season before the 2018 image was taken, in 2017, was wetter than average, while I was unable to determine the rain fall that preceded the image in 2007. As such, it is possible that the 2018 Landsat 8 image shows a non-normal vegetative pattern — or it’s even possible that the 2007 image is showing a non-normal vegetative pattern! I require some more investigation into the historical meteorology of the area before I can say either way. Regardless, I feel that this is a useful first step in investing how LULC change relates to the establishment of ASGM.
Grant, this is an interesting outcome. I did not see a definitions of which lines in your Fig. 3 represented loss vs. gain, but assuming the blue line represents gain, and based on your later comments and our conversation, I gather that the overall change between 2007 and 2018 was positive. You could account/compensate for this by picking a break point other than zero to define “gain” and “loss” – simply use the median change for the whole image, and repeat the analysis. Then your lines will show the amounts of “greater than average gain” and “greater than average loss” as a function of distance from the villages. I wonder if you can get this fixed to include in your final project. Also in your final project I would like to hear what you expect to find and what you actually found: most work suggests that deforestation proceeds radially from villages, so what shape would you expect for the red curve representing the spatial pattern of “greater than average vegetation loss” with distance from mining towns? In addition, it was unclear to me whether the curves differed between farming and mining towns – what would you expect and why? What did you find?
Grant, nice work. As we discussed, the later of your two images appears to be greener overall, producing strong patterning of positive change along riparian corridors, for example. Such corridors are spatially autocorrelated, which might account for the very high positive values of Moran’s I – but it would be important to run Moran’s I for selected different distances classes to explore this further (if you were interested). For Ex 2, I look forward to being able to see your figures when we create more space on the blog. You will be doing this analysis based on images that have been adjusted, as we discussed, correct? Nearly 50% of pixels showed a decline in vegetation cover from 2007 to 2018 – is that correct? If so, and these pixels showing declines are clustered near mining settlements, it’s a good start. At this point, it would be interesting to test whether the spatial patterns of vegetation loss differ among mining settlements of different ages or sizes, and/or how they compare to villages which are not mining settlements.