AI4SmallFarms: Sentinel-1 Extension for Crop Field Delineation in Southeast Asian Smallholder Farms

Crop field boundaries play a central role in agricultural monitoring, yet accurate boundary mapping from satellite imagery remains challenging because performance is strongly affected by field size, fragmentation, and crop phenology. This record extends the publicly available AI4SmallFarms benchmark by adding analysis-ready, multi-temporal Sentinel-1 VH backscatter data coregistered with the existing Sentinel-2 dataset. The original benchmark contains more than 400,000 field polygons across 62 tiles of approximately 5 × 5 km in Vietnam and Cambodia. The Sentinel-1 extension was developed to support research on multimodal crop field boundary mapping in smallholder farming systems. In particular, it enables the joint use of multi-temporal Sentinel-1 SAR data and Sentinel-2 monthly composites for field delineation and boundary refinement tasks. The extension is designed for benchmarking methods that combine complementary radar and optical information, especially in fragmented agricultural landscapes where accurate boundary extraction is difficult. This dataset was used in a study proposing an edge-aware multimodal crop boundary mapping approach based on a dual-stream U-Net with scSE modules and mid-level feature fusion. Boundary refinement was further supported by a composite loss combining binary cross-entropy, Tversky loss, and a Sobel edge penalty. Experiments on the AI4SmallFarms benchmark showed that the multimodal approach produced sharper and more continuous boundaries with fewer spurious edges, particularly for smallholder farms in highly fragmented agricultural areas. The Sentinel-1 data in this record are coregistered to the existing AI4SmallFarms Sentinel-2 benchmark and are intended to be cited together with the original dataset: (https://doi.org/10.17026/DANS-XY6-NGG6).