Inference smoothing may overestimate class probabilities at upper-right corner

[remtav]

At inference, our smoothing window multiplies raw chip probability maps to set higher confidence to predictions in center of chip. As a result of the convolution with a high overlap (low stride) and smoothing on original image, it seems the smoothing may overestimate class probabilities at upper-right corner as is shown by this heatmap on which a water extraction never "activates" (i.e. predicts with constant probability the absence of waterbody). This should be tested thoroughly and independently of the action of any model (i.e. with a numpy array filled with close-to-zero values).