DeepLabV3+ model for high-accuracy binary image segmentation, trained to detect roofs (foreground vs background) and output a grayscale mask.