SWD-Net: A Benchmark Dataset for Small Water Body Detection From Very High-Resolution Imagery

Small and seasonal water bodies store a substantial portion of inland surface water and play a key role in local hydrology, ecosystem resilience, and agricultural sustainability. Yet their small size, irregular geometry, and frequent seasonal variability hinder accurate detection in global water products derived from coarse-resolution imagery. To address this limitation, we present Small Water Body Detection (SWD-Net), a high-resolution, expert-annotated benchmark dataset for small water-body detection from aerial and satellite imagery. SWD-Net consists of 10,085 image–mask pairs (1,024 × 1,024 pixels) sourced primarily from NAIP, Google Earth, and LoveDA, spanning diverse climatic and land-cover conditions. Baseline experiments using five representative segmentation architectures, including U-Net, Mamba U-Net, ViT U-Net, SegFormer U-Net, and Swin U-Net, demonstrate clear architectural contrasts, with transformer-based networks achieving the most accurate delineation of fine-scale and fragmented ponds. Among them, Swin U-Net achieved the best overall performance, with an IoU of 93.53% and an F1-score of 96.56% on the test set. All data, pretrained weights, and evaluation codes are publicly released to facilitate transparent and reproducible research in fine-scale hydrological mapping, and seasonal water monitoring.