Semantic Segmentation of RGB Satellite Urban Images Based on Improved U-Net

With the increase of satellite data in recent years, we can perceive the detailed structure of the Earth’s surface from it, and obtaining this information opens up new directions for remote sensing image analysis. Generating high-quality building segmentation from RGB satellite images can well predict the geometry of cities, while significant progress has been made in semantic segmentation of satellite images based on deep learning, but most existing methods tend to have poor predictions on segmentation boundaries and poor separation of targets in contact with each other in the same category. In this paper, we propose an improved network based on U-Net by incorporating the SimAM parameter-free attention mechanism. In addition, we solve the problem of irregular contours of building segmentation in satellite images by generating weight mappings with a priori information to be added to the calculation of the loss function. We use a competition dataset named Building Missing Maps with Machine Learning on Aicrowed for training and prediction, and demonstrate that our method has good prediction results on the contours of building semantic segmentation independent of the size of satellite images and the number of datasets.