Collaborative Research: A Physics-Informed Flood Early Warning System for Agricultural Watersheds with Explainable Deep Learning and Process-Based Modeling

这十年来，全球洪水和极端降雨事件激增了50%以上，现在发生的频率是1980年的四倍。然而，物理模型在空间尺度上预测洪水事件的能力仍然有限，特别是在像美国中西部这样集约化管理的农业系统中。极端降水和高流量事件观测到的季节模式之间的明显差异，给单独使用降水来预测洪水的发生和严重程度带来了挑战。该项目解决了水文科学中的一个基本问题：流域特征和内陆管理实践如何调节农业主导的流域的降水-径流关系？本项目的建模框架将把流域特征、人类土地利用和管理实践的复杂影响整合到水文预测中。将开发一个早期预警系统，在一个管理系统中预测洪水的发生，并将共享以进一步评估洪水预报性能和不确定性评估。这项研究的首要目标是开发一个数据驱动的、基于物理的早期预警系统，以预测洪水的发生，并支持美国中西部以农业为主的流域的社区。本项目将开发基于图的变压器深度学习方法与基于过程的水文生态建模相结合，以提高洪水预测精度并保持可解释结构。该项目的结果将被测试、共享，并作为一个基于网络的平台上的实时预测工具进行部署，该平台集成了地图功能、高级可视化和移动访问。预警系统将提供给多个用户，特别是代表性不足的社区，以了解洪水对生命和财产的直接影响以及对社区粮食安全、经济和生计的间接影响。该项目由水文科学、刺激竞争研究的既定计划（EPSCoR）和地球科学理事会共同资助，以支持地球科学领域的人工智能/机器学习进步。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Global floods and extreme rainfall events have surged by more than 50% this decade and are now occurring at a rate four times higher than in 1980. However, the capability of physical models in predicting flood events remains limited across spatial scales, especially in intensively managed agricultural systems like the Midwestern U.S. The apparent disparity between observed seasonal patterns of extreme precipitation and high streamflow events presents a challenge when using precipitation alone to predict flood occurrence and severity. This project addresses a fundamental question in hydrologic science: how do watershed characteristics and in-land management practices regulate the precipitation-runoff relationship across agriculture-dominated watersheds? The modeling framework in this project will integrate the complex impacts of watershed characteristics, human land use, and management practices into hydrological prediction. An early warning system will be developed for projecting flood occurrence at a granular level in a managed system and will be shared for further evaluation of the flood forecasting performance and uncertainty assessment.The overarching goal of the research is to develop a data-driven, physics-informed early warning system to predict flood occurrence and support communities in agriculture-dominated watersheds across the Midwestern United States. This project will develop a graph-based transformer deep learning approach integrated with process-based hydro-ecological modeling to improve flood prediction accuracy and keep the interpretable structure. The results of the project will be tested, shared, and deployed as a real-time prediction tool on a web-based platform that integrates mapping capabilities, advanced visualizations, and mobile access. The early warning system will be accessible to multiple users, especially underrepresented communities, concerning the direct impacts of flooding on life and property and the indirect effects on the food security, economy, and livelihood of the communities.This project is jointly funded by Hydrologic Sciences, the Established Program to Stimulate Competitive Research (EPSCoR), and the Directorate for Geosciences to support AI/ML advancement in the geosciences.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.