Background. The solution of design problems in the field of territorial systems management is based on the need for automated analysis of large arrays of geospatial data, including space imagery materials. The purpose of the article is to study and develop effective methods for the semantic segmentation of images containing geospatial data. Methods and materials. The paper studies and uses methods and algorithms for increasing the accuracy and speed of the neural network of the U-Net architecture. Results. Comprehensive studies have been carried out on configuring the U-Net convolutional neural network, in particular, image restoration by nearest neighbor interpolation, adding thinning layers to the architecture. Conclusions. The presented approaches for configuring the U-Net neural network made it possible to increase the accuracy of recognition by the Jaccard metric by 2-3% and increase the speed by 2 times.
Keywords: artificial neural networks, machine learning, deep learning, convolutional neural networks, u-net, semantic segmentation, pattern recognition, geospatial data