As seen in Chart 1. below, the best time for tree mapping in Rwanda is from around May and June to August. Rwanda’s climate is composed of two rainy seasons (February to May and September to December) and two dry seasons in between. Nevertheless, the dry season from June to August is more intense than the other one. Since less rain means there are less clouds, there are better chances of having good atmospheric conditions to obtain satellite data with little cloud cover during that time.
source: WORLD METEOROLOGICAL ORGANIZATION
Adapting to a class schema
Previously, I had four classes: forest, green fields, bare ground and water. To fit with the schema, I decided to keep the forest, bare ground and water classes, and split the green fields class into two categories: grassland and cropland. This left me with 5 categories shown in the gallery below: natural forest, grassland, cropland, bare ground and open water. I did not include more categories because they were either not applicable, or they would not cause confusion. Again, I decided to include water, as there are multiple green water bodies that may have caused classification errors.
Bare Soil 
Water 
Grassland 
Cropland 
Forest
About the mapathon dataset
A key advantage of the mapathon data set is that it allowed us to collect a significant amount of truth points in a very short time period, and most importantly without having to be on the ground. Nonetheless, the data for the different locations was not always from the same time period, so that introduces uncertainty to our classification as these truth points were not from the same dates as our classification data.
Comparison of Landsat-8 and Sentinel-2 outputs
Visual inspection of both classifications suggests that both tree cover maps were very effective at capturing the presence of trees in the landscape. The differences between them were most visible in how clustered the classification was. For example, for the forest patched shown in Figure 1.,using the L8 data produced a tree cover map where forested areas were more compact than with S2 data. As you can see in Figure 2.
The classification was more clustered than with the S2 data shown in Figure 2. For example, for the forest patched shown in Figure 1.,using the L8 data produced a tree cover map where forested areas were more compact than with S2 data.
Figure 1. Satellite image of a forest patch.
Figure 2. Landsat-8 classification of forest patch 
Figure 3. Sentinel-2 classification of forest patch
This could either be attributed to the higher resolution of Sentinel-2 data, or to a lack of data in the gap-filled image. Inspecting the maps, I found that in many areas where Sentinel-2 does not classify tree cover, but Landsat-8 does, there are holes in the Sentinel-2 gap-filled image. Take for example the Sentinel-2 tree cover in Figure 4. In this map, the areas that were not picked up by Sentinel-2 coincide with poorly masked and gap filled areas (clouds) shown in Figure 5.
Figure 4. Example of Sentinel-2 classification of a forest patch 
Figure 5. Sentinel-2 classification overlapping gap-filled images. White areas are poorly masked and gap-filled (clouds)
Regarding the confusion matrices, I think that maybe our classification could not be compared to the truth points that we collected from Dr. Arakwiye’s tool. My reasoning is that those truth points are confirming the presence of trees at the centroid of the sampling area even though there may not be trees in that specific point. Therefore, this could have led to issues where my classifier did not classify that point as tree cover, but the truth point reported that it was. As seen in Tables 1 and 2, there is very poor agreement between my classification maps and the reference points (kappa coefficient < 0.2 in both cases), even though visual inspection suggests that this level of disagreement may not be accurate at all.
