Classification of land based on vegetation cover using deep learning methods for satellite image segmentation

Abstract

This paper examines the application of deep learning methods for the automatic segmentation of satellite images with the aim of classifying land cover types based on their degree of vegetation cover. Three-channel satellite images and vector polygons annotated by experts were used as input data, including classes such as fields, fallow land, various types of tree and shrub vegetation, water bodies, and man-made objects. To solve the segmentation problem, the U-Net and DeepLabv3+ architectures were applied with a pre-trained ResNet50 encoder using combined lossfunctions. As aresult ofthe experiments,theDeepLabv3+modelwiththe ComboLossfunction demonstratedthe bestsegmentation quality metrics. The practical outcome of this work was the implementation of a plugin for the QGIS geographic information system, enabling automatic segmentation and vectorization of satellite data. The proposed approach can be used for land monitoring, landscape change analysis, and decision support in the field of natural resource management.