Fig. 10
From: Using spatial video and deep learning for automated mapping of ground-level context in relief camps

Comparison of model outputs for June and July 2021 against manually digitized shelter locations. In general, the locations and intensities are similar between the maps, especially in map area A, although there is lower intensity in the output for sections of map area B, which is explained by the difference between mapping actual tent locations and creating a smoothed density of visible tents. In the lower map, the high intensity of map area C is captured in both outputs, although there is again some variation in the area immediately to the south. Map area D shows one of the few locations where the manual mapping covers areas not predicted by the model, with one explanation being that there can be more visible extrapolation from the manual mappers if tents are seen to extend beyond what might be recognizable as a tent shape. This area also contains other wooden shelters, and some of these differences are expected