A novel high-resolution satellite imagery approach for surface water detection in the United States

This is the authors' abstract. We don't add key points for this paper.

High-resolution surface-water segmentation supports hydrology, ecological monitoring, and disaster response, but many workflows face trade-offs among spatial resolution, geographic coverage, spectral information, and computational cost. We developed United States Surface-Water (USS-Water), a US-focused red, green, and blue (RGB) dataset containing 1483 Google Earth image tiles at 0.3 m ground sample distance (GSD) from 147 sites in 44 states, with approximately 1.48 billion labeled pixels. Tiles are described by seven water-body/scene categories – creeks, harbors/shipyards, lagoons, lakes, ocean/coastal waters, ponds, and rivers – while the evaluated task is binary water/non-water segmentation. We also propose U-Net+, an efficiency-oriented U-Net variant using patch compression, depth-wise separable convolutions, and lightweight attention. Under the random image-level benchmark, the multiscale residual network (MSResNet-34) achieved the highest intersection over union (IoU) and F1-score, whereas U-Net+ achieved competitive performance (IoU, 83.2%; F1-score, 93.6%) at 6 frames per second (FPS), using 4.2 gigabytes (GB) of graphics processing unit (GPU) memory. Chicago and Gary case studies showed applicability in two urban-industrial scenes. Released materials remain subject to source-imagery licensing.