Development of an algorithm for the Earth remote sensing data classification using deep machine learning methods for analyzing the geosystem model of the territory

The authors propose their solving the task of improving the accuracy of remote sensing data classification under conditions of labeled data scarcity through using a geosystem approach that involves analyzing the genetic uniformity of various-scale territorially adjacent formations and hierarchical levels. The advantage of the proposed GeoSystemNet model is a great number of freedom degrees, which enables flexible configuration of the model based on the task being solved. Testing the GeoSystemNet model for classifying the EuroSAT set, algorithmically expanded from the perspective of the geosystem approach, showed the possibility of increasing the classification accuracy under the conditions of training data scarcity within 9 %, as well as approaching the accuracy of the deep ResNet50 and GoogleNet models. The authors note that the use of the geosystem approach according to the methodology proposed in the article for solving the above-mentioned problem requires an individual project approach to the formation of the data for analysis.