FloodCastBench：面向洪水建模与预测的大规模数据集及基础模型基准

FloodCastBench 是专为机器学习驱动的洪水建模与预报而设计的综合性数据集，涵盖 2015 年英国、2019 年莫桑比克、2022 年巴基斯坦及 2022 年澳大利亚四大洪水事件。数据集分为三大模块：低精度洪水预报（空间分辨率 480 m，含巴基斯坦与莫桑比克事件）、高精度洪水预报（空间分辨率 30 m 与 60 m，含澳大利亚与英国事件）及辅助数据（DEM、土地利用/覆盖、降雨时序、地理参考与初始条件）。所有时空数据均以 TIFF 格式存储，时间分辨率为 300 s，文件按时间步长递增编号。数据集结合有限差分数值解，对洪水动态过程进行高分辨率（30 m × 300 s）模拟，并通过 SAR 测图与地面勘测轮廓进行参数校准与结果验证。此外，FloodCastBench 还提供神经网络洪水预报基础模型基准，支持跨区域、时空及降尺度洪水预报研究，为机器学习在水文灾害领域的应用提供了标准化、可重复的测试平台。

FloodCastBench is a comprehensive dataset designed for machine learning-driven flood modeling and forecasting, covering four major flood events: the 2015 event in the United Kingdom, the 2019 event in Mozambique, the 2022 event in Pakistan, and the 2022 event in Australia. The dataset is divided into three modules: low spatial resolution flood forecasts (with a spatial resolution of 480 m, covering the Pakistan and Mozambique events), high spatial resolution flood forecasts (with spatial resolutions of 30 m and 60 m, covering the Australia and United Kingdom events), and auxiliary data (including DEM, land use/cover, rainfall time series, georeferenced data and initial conditions). All spatiotemporal data are stored in TIFF format, with a temporal resolution of 300 s, and the files are numbered incrementally according to time steps. The dataset integrates finite difference numerical solutions to simulate flood dynamic processes at a high resolution of 30 m × 300 s, and conducts parameter calibration and result validation using SAR mapping and ground survey profiles. Furthermore, FloodCastBench also provides baseline neural network flood forecasting models, supporting cross-regional, spatiotemporal and downscaling flood forecasting research, and offering a standardized and reproducible test platform for the application of machine learning in the field of hydrological disasters.

搜集汇总

背景与挑战

背景概述

FloodCastBench是一个面向洪水建模与预测的大规模数据集，包含四大洪水事件的高、低精度预测数据及辅助数据，支持跨区域、时空和下尺度的洪水预测研究，为机器学习在水文灾害领域的应用提供了标准化测试平台。