A Deep Learning Approach to Distributed Rainfall-Runoff Modeling with Spatiotemporal Interpretation.

Distributed rainfall-runoff modeling is crucial for understanding the complex process that integrates hydrological cycles, contaminates transport, and land-use/land-cover change. An accurate model typically exhibits tremendous temporal and spatial heterogeneity. Process-based distributed models embed solid physical mechanisms and detailed hydrological processes but suffer from region-specific parameterization and time-consuming calibration. Data-driven conceptual models have been widely applied but mostly in a lumped format that disregards spatial heterogeneity without the inclusion of comprehensive geophysical parameters and have been criticized for lack of proper physical mechanism. In this study, we adopted a computer-vision-based deep learning architecture, ViT (Vision Transformer), for distributed rainfall-runoff simulation where both gridded data time series and station records can be adopted. Apart from forecasting, our model could offer interpretation that links spatiotemporal inputs and forecast outcomes via tracking gradient weighted attention score through the Transformer network. The case study was examined in three subbasins of the Trinity River Basin in Texas. Our results demonstrated satisfactory forecasting performance and spatiotemporal interpretation, which could support scenario planning and decision making.