This article presents a method for classifying multi-time multispectral images of the Earth's surface using the convolutional deep learning neural network U-net. Images of visible and infrared wavelengths were obtained using a multispectral optoelectronic system of an unmanned aerial vehicle and were used to construct orthophotoplanes of the terrain. Based on the data obtained, a neural network was trained to solve the problems of detecting man-made objects. The method of intelligent recognition of remote monitoring objects, based on deep learning and assessments of thermophysical parameters, allows you to create a phono-target environment using a genetic algorithm. This algorithm solves the coefficient inverse problem of thermal conductivity and provides estimates of the thermophysical parameters of materials. To train the model, 18 classes of objects were introduced, which were studied based on the difference in thermal contrast between man-made objects and the background (anthropogenic or natural landscape). The survey of the earth's surface was carried out 6 times during the day with an interval of 4 hours. The experiment was conducted in the summer of 2021, on specific dates of August 4-5. In the tasks of detecting and classifying man-made objects, it was found that the model demonstrates applicability with varying reliability. The conducted research shows that during the operation of the model, the desired classes of objects were discovered.

Deep learning, classification, segmentation, remote monitoring neural network, genetic algorithm, phono-target environment, optoelectronic systems, thermophysical parameters

Gromov Yuri Yurievich,  ,  orcid.org/0000-0003-3313-2731,  Tambov State Technical University

Ishchuk Igor Nikolaevich,  ,  orcid.org/0000-0002-8140-3746,  Tambov State Technical University

Rodionov V.V.,  ,  orcid.org/0009-0004-6327-8082,  Tambov State Technical University