Mapping of agricultural land based on neural networks using sentinel-2 and landsat-8 data

This study presents a detailed comparative analysis of the performance of advanced algorithms across three neural network models for mapping agricultural lands and classifying land cover based on multispectral satellite imagery from Sentinel-2 and Landsat-8, acquired for the Berestyn district of Kharkiv region. The primary objective was to determine the optimal methodology for automated processing and interpretation of remote sensing data for agricultural land mapping, with further applications in monitoring, environmental assessment, and land use planning. Within the study, land cover was classified into four categories: vegetation, agricultural lands, water bodies, and built-up areas. Three classification approaches were applied: a Deep Neural Network (DNN), a Convolutional Neural Network (CNN), and the Random Forest (RF) algorithm. The DNN combines mechanisms for extracting both spectral and spatial features, allowing for efficient consideration of local textural characteristics of objects as well as their spectral differences. Experimental results confirmed its superiority over the other methods: classification accuracy reached 95,90 % for Sentinel-2 and 90,90 % for Landsat-8. By comparison, CNN achieved 93,8 % and 87,5 %, while RF attained 90,3 % and 87,2 %, respectively. The advantage of DNN was particularly evident in the analysis of areas with complex mosaic land use structures, where traditional algorithms often misinterpret class boundaries. At the same time, certain limitations were identified: high computational complexity and substantial memory requirements make it less suitable for real-time analysis in resource-constrained or field conditions. The practical value of the obtained results lies in the potential to implement this methodology for automated monitoring of crop conditions, early detection of stress factors and plant diseases, assessment of crop growth dynamics, and control of land use changes under the influence of anthropogenic and natural factors. Future research will focus on optimizing the DNN architecture to reduce resource consumption without compromising accuracy, integrating additional data types, including hyperspectral and radar imagery, and developing hybrid solutions capable of combining the advantages of deep learning with classical image processing methods to enhance the versatility and adaptability of remote monitoring systems.