AI Image Techniques Forecast Hourly Flash Floods Across U.S. River Basins

Researchers at Penn State and the Scripps Institution of Oceanography have developed a novel artificial intelligence system that significantly improves hourly flash flood forecasting across the United States. Led by civil and environmental engineering professor Chaopeng Shen, the team published their findings in Water Resources Research, demonstrating how generative AI techniques traditionally used for image synthesis can be adapted to predict rapidly evolving hydrological events. Flash floods account for the vast majority of flood-related disasters globally, often causing severe damage within hours. Traditional daily forecasting models frequently miss critical peak water levels because they fail to capture the rapid, hour-by-hour fluctuations that characterize these events. To address this limitation, the researchers applied diffusion models, a class of generative AI originally designed for image creation. Unlike standard large language models, diffusion training involves feeding the system complete datasets, introducing controlled noise, and teaching the algorithm to reconstruct accurate patterns from that noise. This approach proves highly effective for managing the erratic nature of rainfall and streamflow data. The model was trained on hourly streamflow records from 516 continental U.S. river basins spanning 1990 to 2003 and validated against an unseen 2009 to 2014 dataset. Testing revealed that the diffusion-based system outperforms existing hourly deep-learning forecasting models, particularly in capturing high-flow periods. The architecture also features a dynamic inpainting capability, allowing meteorologists to inject recent, near-real-time gauge observations directly into the model without requiring full retraining. This reduces predictive uncertainty and enables continuous refinement as new weather data arrives. Additionally, the system can operate in reverse, estimating hourly rainfall volumes based on recorded streamflow at river gauges. Performance mapping using the Nash-Sutcliffe Efficiency metric confirms substantial accuracy improvements across a broad geographic range. The research team emphasizes that the technology is not intended to replace conventional hydrological methods but to complement them. Future iterations will explore hybrid training frameworks that integrate additional physical variables to further enhance reliability. Ultimately, the developers aim to deploy a robust, operationally viable forecasting platform capable of delivering timely warnings that protect vulnerable communities from sudden water surges.