My supervised classification map that I compared to Hansen’s tree cover map identified five classes (water in light blue, urban in orange, agriculture in purple, forest in green, bare ground/other in yellow). From visual comparison between my map and satellite imagery, my classification map does a pretty good job identifying distinct features like urban areas and water bodies but does a poor job separating regions with mixed land cover. In particular, the classification mixes up agriculture and bare land, which I anticipated when creating these classes because the two can look similar depending on crop productivity or non-agricultural brush cover. Water and urban areas were also misidentified, but this occurred most often in areas with hills or trees creating shadows which affect the signature of the image, such as in Figure 1.

On the scale of the whole image in Figure 2, my forest classification was not very accurate when compared with Hansen’s map. Comparing Hansen’s data of forest at a 30% threshold, all of the pixels in the image were classified as either false positive (I identified as forest and Hansen’s did not) or true negative (both Hansen’s and I identified as not forest). My classifier correctly identified all of the forests but included additional areas without forest in an error of commission.

It is difficult to identify vegetation height and land use from a satellite image, which are both factors that Hansen’s data use to define a forest that I was not able to glean from the image. Figure 3 shows that pixels that I chose for my forest cover training points were not considered to be forest by Hansen’s. Specifically, the area outlined below is the Foret Classee de Yamba Berete, a forest reserve in Chad, that I used to train my classifier. However, Hansen’s does not classify this as forest, and thus there is disagreement.

My classifier had a higher rate of agreement with Hansen’s within the study area, as shown on the agreement map. Some of the features that we disagree upon are the lines of rivers, which are visible in Figure 4. Dense vegetation grows along rivers in this area due the dryness of the landscape, so these clusters of vegetation along the river that appear to be forest by my classification would not be considered full forest by Hansen’s definition.

If my tree map within the study area were the true reflection of what is on the ground, the most accurate representation of Hansen’s map would have a threshold of 0.1, meaning that a pixel only has to contain at least 10% forest cover for the pixel to be considered forest. Table A shows the True Positive Rate (TPR) and True Negative Rates (TNR) for four different forest cover thresholds. Thresholds above 0.3 remained the same as the classification values at 0.3. The TPR is 0 for both 0.2 and 0.3 because neither threshold had true positive values. Of the thresholds tested (including those not shown on this table), 0.1 had the highest balanced rate and the highest possible TPR. 10% is a relatively low threshold for forest cover in a pixel (especially considering the default for Hansen’s is 30%). As a sub-Saharan landscape, the study area is very dry and does not have large swaths of forest (the vegetation appears scattered), so to classify anything as forest requires a low threshold for what should be considered at all.

	Threshold = .05	Threshold = .1	Threshold = .2	Threshold = .3
TPR	0.6195877	1.0000000	0.0000000	0.0000000
TNR	0.8790142	0.8431880	0.5008536	0.8354294
Balanced	0.7493009	0.9215940	–	–