Using deep learning algorithms for texture segmentation of ultra-high resolution satellite images

Abstract

This paper presents the results of textural segmentation of satellite images with spatial resolution <1 m using U-Net convolutional neural networks. To conduct numerical experiments, a panchromatic image of the WorldView-2 test site on the territory of the Bronnitsky Forestry (Moscow region) used. The possibilities of automating the selection of neural network parameters based on genetic algorithms investigated. The proposed method makes it possible to effectively segment the main types of natural and man-made objects, as well as to distinguish structural classes of woodlands.