GEOINFORMATION TECHNOLOGY NEURAL NETWORK SEGMENTATION FOR LAND COVER MAPPING

Abstract

The relevance of developing modern technologies for land cover segmentation is growing due to increased requirements for accurate monitoring and management of land resources including for agricultural purposes. Traditional segmentation methods often lack accuracy in classifying complex classes such as crops, trees, buildings, and roads. The work aims to develop geographic information technology for extracting multiple features from Sentinel-2 satellite images and using them to segment the land cover using the ResNet neural network. Methodology. This study uses Sentinel-2 images for land cover analysis. First, the images undergo preprocessing, which includes atmospheric correction and geometric and radiometric calibration. Then, data is normalized to improve the stability of the neural network training. At the next stage, the images are processed to extract spectral, morphological, and textural features, which are the input to the ResNet model. The model uses convolutional layers and the ReLU activation function to extract features automatically. A fully connected layer with Softmax and Cross-Entropy functions is used for classification. After training, the model classifies each pixel, creating a segmented image that shows different classes of land cover, including farmland, buildings, trees, and roads. The scientific novelty of the research is the development of the latest methodology for processing Sentinel-2 satellite images, including integration of complex pre-processing, data normalization, multimodal feature extraction, and the use of deep neural networks for automatic feature extraction and classification. The new approaches to atmospheric, geometric, and radiometric correction, as well as the use of ResNet with ReLU activation and fully connected layers with Softmax and Cross-Entropy functions, improve the accuracy of classification and detail of land cover segmentation. Conclusions. The study showed that the proposed technology provides a significant improvement in classification accuracy and quality compared to traditional methods such as IsoData, K-means, SVM, Minimum Distance, Maximum Likelihood, and Parallelepiped. The results demonstrate that the ResNet-based technology demonstrates high precision in segmenting the main land cover classes–crops, trees, buildings, and roads–which is crucial for effective land monitoring and management.