Aiming at the problem that it is difficult to accurately extract the building roof area in the assessment of rooftop photovoltaic resources based on satellite images, this paper proposes a method for the assessment of rooftop photovoltaic resources on multi-building roofs based on the improved Mask-RCNN neural network. First, in order to improve the data quality of the model training samples, this paper produces 5000 multi-building roof image datasets based on Google Maps and divides the roof images into five types. Then, based on the original Mask-RCNN network, this paper improves the feature pyramid network by adding bottom-up fusion paths to improve the feature learning and extraction ability of the network, and introduces an image classification network to reduce the impact of multiple types of roofs on the network extraction performance. Finally, based on the constructed image dataset, the effectiveness and accuracy of the method proposed in this paper are verified.