STURM-Flood: a curated dataset for deep learning-based flood extent mapping leveraging Sentinel-1 and Sentinel-2 imagery

Flooding is a major global natural disaster exacerbated by climate change and urbanization. Timely assessment and mapping of inundations are crucial for preventive and emergency measures, driving the demand for curated global geospatial data to implement novel algorithms. This study introduces the STURM-Flood dataset, a high-quality, open-access, and DL-ready resource for flood extent mapping using Sentinel-1 and Sentinel-2 satellite imagery, combined with ground-truth data from the Copernicus Emergency Management Service. The dataset encompasses 21,602 Sentinel-1 tiles and 2,675 Sentinel-2 tiles, each measuring 128 × 128 pixels at 10 m resolution, alongside corresponding water masks covering 60 flood events globally. Two U-Net models evaluated the dataset: Sentinel-1 achieved 83.61% test accuracy and 0.8327 weighted F1-score, while Sentinel-2 yielded 92.75% test accuracy and 0.9243 weighted F1-score. These results underscore the dataset’s potential in developing robust models for water extent mapping. STURM-Flood dataset aims to provide a valuable resource for research and development in flood mapping and disaster management. Future research could focus on expanding and refining different approaches and data sources for broader applications. The reference data and code are openly available in the Zenodo repository https://doi.org/10.5281/zenodo.12748983 and GitHub repository https://github.com/STURM-WEO/STURM-Flood.

洪涝是全球主要自然灾害之一，其影响因气候变化与城市化进程而加剧。及时开展洪涝灾情评估与范围制图，对于防灾与应急处置工作至关重要，这也催生了对经过筛选整理的全球地理空间数据的需求，以支撑新型算法的研发与落地。本研究推出STURM-Flood数据集，这是一款高质量、可开放获取且适配深度学习（Deep Learning）的资源，可依托哨兵1号（Sentinel-1）与哨兵2号（Sentinel-2）卫星影像开展洪涝范围制图，并结合哥白尼应急管理服务（Copernicus Emergency Management Service）提供的地面真值数据。该数据集包含21602幅哨兵1号瓦片与2675幅哨兵2号瓦片，所有瓦片均为10米分辨率、128×128像素规格，同时附带对应全球60起洪涝事件的水体掩膜数据。研究采用两款U-Net模型对该数据集进行性能验证：基于哨兵1号影像的模型测试准确率达83.61%，加权F1分数为0.8327；基于哨兵2号影像的模型测试准确率达92.75%，加权F1分数为0.9243。上述结果充分证明了该数据集在研发高精度水体范围制图模型方面的应用潜力。STURM-Flood数据集旨在为洪涝制图与灾害管理领域的研发工作提供宝贵的数据资源。未来研究可聚焦于拓展与优化各类方法及数据源，以实现更广泛的应用场景。相关参考数据与代码已公开上传至Zenodo知识库（https://doi.org/10.5281/zenodo.12748983）与GitHub仓库（https://github.com/STURM-WEO/STURM-Flood）。

搜集汇总

背景与挑战

背景概述

STURM-Flood是一个高质量的开放数据集，专为深度学习洪水范围映射设计，包含大量Sentinel-1和Sentinel-2卫星图像及对应水掩膜，覆盖全球60个洪水事件。数据集经过U-Net模型评估，显示出较高的准确率，适用于洪水映射和灾害管理研究。