From visual inspection, my classification map is not fantastic comparing to Hansen’s tree cover map (where tree cover is only considered when greater than 30%). In many places, my map said there were trees where the Hansen image said there were not (see red pixels in Figure 1; Table 1). 

However, looking at the area chart, there are more pixels that agree than disagree (classes 1 + 4 > 2 + 3;  Figure 2). As the threshold for what is considered tree cover in Hansen’s classification becomes more stringent (increases), the number of pixels in classes 1 and 3 decreases, while the number of pixels in classes 2 and 4 increase (Figure 2-3).  

There is most agreement about tree cover (class 1) between my map and Hansen’s map where Hansen’s tree cover threshold is set to 10%. However, in this case there is very low agreement about non-tree cover, as so much of Hansen’s image is considered tree. When Hansen’s tree cover threshold is 25%, there is a little less agreement about tree cover (although it is still quite high) and a relatively high agreement about non-tree cover; however, there are many pixels in class 2, meaning there were many trees in my map where there were not in Hansen’s map.  Although the balanced accuracy rate increases with threshold stringency, this is due to the increase in true negatives; true positives are lowest at the highest threshold.

Therefore, when looking solely at tree cover agreement, the lowest threshold yields the highest true positives. However, accepting this as the “true” tree cover is tricky due to the high rate of false positives that comes with this assumption. I would therefore say that at a threshold of approximately 35%, there is the most tree cover accuracy because true positives are relatively high while false positives are relatively low (Figure 4).