Extending spectral indices from multispectral satellite data using U-Net segmentation

作者：A. Gayibov, V. Gasimov · 发表于：RADIOELECTRONIC AND COMPUTER SYSTEMS · 年份：2026 · DOI：10.32620/reks.2026.1.14 · 被引用次数：2

This study focuses on developing a unified, reproducible cloud-to-model pipeline for parcel-scale cropland delineation from multispectral Sentinel-2 and Landsat imagery in the Kur–Araz region. The goal of this study is to produce accurate, boundary-faithful, and computationally efficient farmland maps that remain transparent, scalable, and deployable on commodity hardware for use by government agencies, water authorities, and agricultural producers. The tasks to be addressed include the following: specification of the study scope, data sources, and evaluation protocol that integrates pixel-wise and boundary-sensitive accuracy metrics; construction of a multi-index feature stack in addition to surface reflectance bands, followed by screening of features for cross-seasonal stability; design and training of a memory-efficient U-Net architecture with a hybrid loss that simultaneously balances calibration and overlap; and validation of model generalization across different cropping seasons and neighboring subregions of Kur–Araz, with complete provenance tracked from preprocessing through evaluation. The methods used in the study include cloud-native preprocessing in Google Earth Engine, consisting of cloud masking, seasonal compositing, medoid and percentile mosaicking, and stratified patch sampling with spatial blocking. The datasets were exported to TFRecords and used to train a compact U-Net encoder–decoder network with skip connections and a hybrid objective function. Training...