Multisource Merging of Daily Precipitation Products over the U.S. Southwest Using Spatiotemporal Graph Representation Learning
Omid Zandi1, Rozhin Yasaei2, Ali Behrangi1
1Department of Hydrology and Atmospheric Sciences, The University of Arizona, Tucson, AZ
2 College of Information Science, The University of Arizona, Tucson, AZ
Accurate daily precipitation estimates are essential for effective water resource management. However, no single dataset provides both complete spatial coverage and high local accuracy. Reanalysis products such as ERA5 offer spatially consistent coverage, but they can underestimate localized extremes, particularly during intense convective events like the North American Monsoon. Satellite-based products such as IMERG provide near-global coverage and capture storm structures, yet they rely on indirect observations of cloud properties to retrieve rainfall, which can introduce uncertainties, especially over complex terrain. Rain gauges offer highly accurate point observations, but they are spatially sparse and must be interpolated to produce gridded precipitation maps.
This project proposes a novel Graph Neural Network (GNN) framework for multisource merging of daily precipitation within a unified learning architecture. The model integrates ERA5 and IMERG inputs while explicitly representing spatial relationships between neighboring locations using a graph structure. A Graph Convolutional Network (GCN) captures spatial dependencies across connected locations, while a Recurrent Neural Network (RNN) models temporal evolution, allowing the system to learn how storms develop and propagate over time. By jointly modeling spatial and temporal processes, the framework aims to better represent regional variability and extreme events.
The approach will be tested over the U.S. Southwest, with a focus on Arizona. We will evaluate different training strategies and loss functions and explore multi-task extensions that combine rainfall occurrence classification with intensity regression. The goal is to develop a scalable data-merging framework capable of producing improved daily precipitation maps to support hydrological decision-making.