Abstract
A region-based convolutional neural network image segmentation approach (Mask RCNN) was applied to identification of flat rooftops from satellite imagery in the city of Jeddah in Saudi Arabia. The model was trained on a small sample of rooftops (202) digitized from a 0.5 m resolution image (covering 0.21 km(2)) and then was applied to an independent area 4.5 km away. The precision and recall of the model were 0.98 and 0.96 respectively in terms of identifying rooftops in the independent area. A spatially stratified sample of rooftops was drawn from those identified by the model and the median roof area of the sample was not significantly different from the area as a whole. The results, although at a small scale, demonstrate the effectiveness of this approach for selecting buildings with appropriate rooftops for solar photovoltaic (PV) installation, in the context of closely spaced flat-roofed buildings, without requiring cadastral mapping or LIDAR datasets.