Delineation of Agricultural Field Boundaries Using Machine Learning: Based on the Case of the Zhetysai District, Kazakhstan

Accurate delineation of agricultural field boundaries is essential for agro-production monitoring, irrigation management, and sustainable land use planning. Manual digitization is labour-intensive and poorly scalable, while conventional segmentation algorithms are sensitive to spectral heterogeneity and weak inter-field contrast. The Zhetysai district of southern Kazakhstan – where approximately 45-50% of irrigated farmland exhibits salinization – exemplifies the urgent demand for automated field mapping. This study evaluates the Meta AI Segment Anything Model (SAM) with Low-Rank Adaptation (LoRA) fine-tuning as a parameter-efficient, scalable approach to automated field boundary delineation. High-resolution Esri World Imagery served as the primary data source. A training dataset of 705 manually digitized field polygons was prepared in ArcGIS Pro and exported as classified tiles. SAM – pre-trained on over one billion masks – was fine-tuned via the SAMLoRA architecture, which freezes the Vision Transformer encoder and trains only lightweight adapter layers, enabling efficient domain adaptation from limited annotations. The model was trained on a CUDA GPU with early stopping, then applied district-wide to generate a pixel classification raster, which was vectorized into closed field polygons and validated through expert visual inspection. Fine-tuning achieved 91% overall pixel accuracy on the validation subset. Applied to the full study area, the model delineated 4,060 field polygons covering 809.76 km2 . Expert verification confirmed high boundary correspondence and negligible false positives. The results demonstrate that a foundation model can be effectively adapted to a new agrolandscape with minimal labelled data, yielding a reproducible and scalable geospatial workflow suitable for precision irrigation management, sown area assessment, and land resource governance across irrigated regions.