Modification of the convolutional neural network architecture for determining the category of a land plot from satellite images

Correct classification of land plots by their types, for example, such as forest, agricultural, urbanized, water bodies, and others, is relevant for remote sensing of the Earth and the development of geoinformation technologies. The accuracy and reliability of the results of such categorization are of paramount importance for the efficient use of natural resources, rational land use, and environmental monitoring. The article presents an approach to solving the problem of categorizing land plots based on satellite images by applying a modified standard model of a convolutional neural network. The main attention is paid to the modification of the network architecture in order to improve the accuracy of land plot classification. The authors propose an approach to training and optimizing the network in order to solve this problem. The stages of data preparation are discussed in detail, including preprocessing satellite images, annotating them, and creating high-quality training samples. The presented approaches to network training and optimization include the use of modern regularization techniques, adaptive learning methods, and class balancing strategies, which allows efficient processing of both large amounts of data and more limited sets of specific information. To test the approach’s operability and obtain the values of quality indicators, experiments were conducted to train and test the model on various sets of satellite image data. The results of the experiment suggest that the accuracy of categorization achieved on the basis of the created model meets the requirements of the Federal Service for State Registration, Cadastre and Cartography for studying remote Territories for the suitability of land for their rational use, and the proposed method can be used to solve practical problems.