Research map / Detection and segmentation / CNN features
GLH-Water: A Large-Scale Dataset for Global Surface Water Detection in Large-Size Very-High-Resolution Satellite Imagery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GLH-Water, a globally distributed dataset of 250 large-size (12,800x12,800) 0.3m VHR optical images with 40.96B labeled water pixels, benchmarks 12 segmentation models, and proposes a pyramid consistency loss (PCL) baseline.
Problem
Detecting global surface water in large-size very-high-resolution satellite imagery lacks suitable public datasets and methods, limiting global high-resolution water mapping and flood applications.
Approach
A large-scale, globally sampled VHR surface water dataset with whole-scene-sized images, plus a pyramid consistency loss that enforces inter-layer (multi-field-of-view) and intra-layer (overlapping tile) consistency to preserve water body continuity.
Key results
- Proposed PCL with PSPNet-Res50 reaches 82.26% IoU on GLH-Water test, +7.07 over PSPNet baseline (75.19%) and +2.4 over next best (DeepLab v3+ 79.80%).
- Model trained on LoveDA drops to 50.13% IoU on GLH-Water test vs 79.80% for model trained on GLH-Water, indicating strong cross-dataset generalization gap.
- In the Yangpu District pilot area, the GLH-Water-trained HRNet-48 achieves 75.99% IoU, outperforming models trained on other datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PCL to PSPNet-Res-50 | PSPNet-Res-50 · GLH-water test set | IoU (%): 75.19 -> 82.26 | Helped |
| Add inter-layer consistency loss (ablation) | PCL with PSPNet-Res-50 · GLH-water test set | IoU (%): +5.89% | Helped |
| Add intra-layer consistency loss (ablation) | PCL with PSPNet-Res-50 · GLH-water test set | IoU (%): +5.50% | Helped |
| Replace the designed consistency loss with vanilla L2 loss (ablation) | PCL with PSPNet-Res-50 · GLH-water test set | IoU (%): 75.19 -> 71.46 | Hurt |
| Remove the second image-pyramid layer (ablation) | PCL with PSPNet-Res-50 · GLH-water test set | IoU (%): 82.26 -> 80.13 | Hurt |
| Remove the third image-pyramid layer (ablation) | PCL with PSPNet-Res-50 · GLH-water test set | IoU (%): 82.26 -> 81.37 | Hurt |
| Add PCL to FCN8s-VGG16 | FCN8s-VGG16 · GLH-water test set | IoU (%): 73.66 -> 75.78 (+2.12) | Helped |
| Add PCL to DeepLabV3+-Res50 | DeepLabV3+-Res50 · GLH-water test set | IoU (%): 79.80 -> 81.33 (+1.53) | Helped |
Surprising findings
- Generic semantic segmentation models generally outperform surface-water-specific detection methods (e.g., MECNet 44.67% IoU).
- Ultra-high resolution methods such as ISDNet (53.04%) and MagNet (62.77%) underperform generic models on this task.
Limitations
- PCL suffers from low inference efficiency (1.34 FPS) and high computational cost.
- Dataset imagery sourced from Google Earth, which may restrict licensing and redistribution (not stated explicitly).
Experimental setup
Datasets: GLH-water LoveDA DeepGlobe
Benchmarks: GLH-Water LoveDA DeepGlobe 2020 GF challenge
Metrics: IoU FPS
Baselines: MECNet MSResNet MSCENet FCN8s PSPNet DeepLab v3+ HRNet-48 STDC-1446 MagNet FCtL ISDNet
Builds on: focal loss DeepLab v3+ PSPNet GradCAM
Closest papers by meaning
- Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation (AAAI 2025 · similarity 0.84)
- Kuro Siwo: 33 billion $m^2$ under the water. A global multi-temporal satellite dataset for rapid flood mapping (NeurIPS 2024 · similarity 0.84)
- SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing (AAAI 2024 · similarity 0.84)
- Landsat30-AU: A Vision-Language Dataset for Australian Landsat Imagery (AAAI 2026 · similarity 0.83)
- Locate Anything on Earth: Advancing Open-Vocabulary Object Detection for Remote Sensing Community (AAAI 2025 · similarity 0.83)
- MMM-RS: A Multi-modal, Multi-GSD, Multi-scene Remote Sensing Dataset and Benchmark for Text-to-Image Generation (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.