METHOD OF PROCESSING IMAGES OF REMOTE EARTH PROBING USING A NEURAL NETWORK WITH A FINE FOR ACCURACY OF THE SEGMENTATION BOUNDARY Russian patent published in 2021 - IPC G06N3/02 G06T7/10 G01C11/04 G06K9/46 

FIELD: data processing.

SUBSTANCE: invention relates to a method of processing geospatial satellite data. Method involves obtaining satellite images and storing them in a database, then shots are normalized, segmented and processed by a neural network, trained by means of final loss function, presented as linear sum of three terms: binary cross-entropy, loss function based on Intersection over Union metric, and a loss function based on a differentiated analogue of the metric of estimating the quality of the boundary F-measure, obtained as follows: obtaining segment boundaries maps, based on which maps of boundaries of greater thickness are obtained, then obtaining indicators of accuracy and completeness, by means of pixel product of obtained segment maps, based on which a differentiable analogue of the quality metric of the boundary F-measure is obtained; at the output of the neural network receiving images, where the value of each pixel is the probability of the pixel belonging to the determined object class; pixel values are rounded off by a threshold value, and the obtained segments are connected to form a segmented satellite image map.

EFFECT: technical result consists in improvement of satellite images processing quality.

4 cl, 4 dwg