Neural Network Approach for Segmentation of Agricultural Lands on Satellite Images

Abstract

Precision mapping and monitoring of agricultural lands using satellite imagery have become crucial for optimizing agricultural practices. This research focuses on ex-ploring the effectiveness of deep learning models, particularly U-Net modifications, for semantic segmentation of agricultural lands in satellite images. Recent Studies and Publications Analysis. Recent advancements in convolutional neural networks (CNNs) have shown promising results in various semantic segmentation tasks, including medical imaging, flood mapping, and environmental monitoring. Studies such as "Residual wave vision U-Net for flood mapping using dual polarization Sentinel-1 SAR imagery" and "Deep learning-based hybrid feature selection for the semantic seg-mentation of crops and weeds" underline the adaptability of U-Net architectures to di-verse data characteristics, motivating their application in agricultural land segmenta-tion. Research Objective. The primary aim of this study is to assess the applicability and efficiency of modified U-Net architectures in accurately segmenting agricultural lands from satellite images. It seeks to identify optimal model modifications that enhance seg-mentation accuracy while maintaining computational efficiency, contributing to more ef-fective agricultural monitoring. Main Body of Research. Utilizing satellite images from the Copernicus HUB archive, this work experiments with various U-Net architectures incorporating residual blocks, normalization methods, and regularization techniques. The study compares the perform-ance of these models in segmenting agricultural lands, highlighting the impact of archi-tectural enhancements on improving model precision and generalization capabilities. Conclusions. The research concludes that specific modifications in the U-Net archi-tecture significantly enhance the segmentation accuracy of agricultural lands in satellite images. Implementing residual blocks, batch normalization, and dropout proved effective in overcoming overfitting, suggesting a promising direction for future work in geospatial data processing for agriculture. Further investigation into hyperparameter tuning, data-set expansion, and ensemble methods is recommended to refine the models' predictive performance.