This is a solar panel segmentation in computer vision. Here are some examples:
Maybe we can try Dice loss:
def dice_loss(score, target):
target = target.float()
smooth = 1e-5
intersect = torch.sum(score * target)
y_sum = torch.sum(target * target)
z_sum = torch.sum(score * score)
loss = (2 * intersect + smooth) / (z_sum + y_sum + smooth)
loss = 1 - loss
return loss