Exploring Data-Driven Models for Compound Flood Forecasting: A comprehensive benchmark

ICLR 2026 Conference Submission19245 Authors

Keywords: Machine Learning, Benchmark, Environmental Science

Abstract: Compound flood forecasting remains challenging due to complex interactions between meteorological, hydrological, and oceanographic factors, a challenge intensified by climate change. Traditional physics-based methods, such as the Hydrologic Engineering Center’s River Analysis System, are often time-inefficient and non-executable due to lack of geographic data. While machine learning shows promise over traditional physics-based methods in both accuracy and efficiency, the lack of comprehensive datasets has hindered systematic evaluation of data-driven approaches. To address this gap, we introduce {\dataset}, a comprehensive benchmark for compound flood forecasting using real-world data from South Florida. Our benchmark integrates four critical factors: tide, rainfall, groundwater, and human management activities, enabling systematic comparison of forecasting methods. We evaluate six modeling paradigms: Multilayer Perceptrons, Convolutional Neural Networks, Recurrent Neural Networks, Graph Neural Networks, Transformers, and Large Language Models. Through extensive experiments, we analyze the impact of key features, temporal dependencies, and spatial relationships on forecasting performance. The varying results across approaches highlight each method's distinct capabilities in capturing compound flood dynamics. By providing this benchmark with open code and data, we aim to accelerate progress in flood forecasting through collaborative research between machine learning and environmental science communities.

Primary Area: datasets and benchmarks

Submission Number: 19245