Neural Network Approach to Segmentation of Economic Infrastructure Objects on High‐Resolution Satellite Images

The problem of semantic segmentation of infrastructure objects in high‐resolution satellite images is considered. This task is an integral part of the method for constructing digital terrain models. Semantic segmentation involves classes such as buildings, roads, and railways. A set of labeled satellite images is collected, and a neural network architecture is selected and trained. In order to reduce the imbalance of classes in the training sample, a probabilistic method of augmentation is developed and applied.