Edge-Aware Crop Field Boundary Mapping From Multimodal Sentinel-1 and Sentinel-2 Data

Crop field boundaries plays a pivotal role in agricultural monitoring. Accurately mapping these boundaries using satellite data, however, remains a significant challenge as the quality of the results is strongly influenced by the field size and the phenological state of the study area. Leveraging the complementary information of Sentinel-1 (S1) and Sentinel-2 (S2) data, this work proposes an edge-aware crop boundary mapping approach that fuses multitemporal S1 with S2 monthly composites in a dual-stream U-Net supported by Spatial and channel squeeze-and-excitation (scSE) and mid-level feature fusion. To refine the boundary delineation, we introduce a composite loss combining binary cross-entropy, Tversky loss, and a Sobel edge penalty. Our experiments demonstrate that the proposed multimodal U-Net architecture achieves the highest accuracy, with the Sobel loss term contributing edge sharpness and continuity. On the publicly available AI4SmallFarms benchmark, which includes over 400,000 field polygons across 62 tiles ( 5×5 km) in Vietnam and Cambodia, our approach produces sharper, continuous boundaries and fewer spurious edges, particularly for smallholder farms with highly fragmented agricultural areas. The benchmark is extended by adding analysis-ready, multi-temporal S1 VH images (can be downloaded from https://doi. org/10.5281/zenodo.18921642) coregistered with the existing S2 data. The source code for the proposed framework is available at https://github.com/mf-celik/multimodal-crop-boundary.