A Hierarchical Deep Learning Framework for Runoff Prediction Using Raster‐Based Spatial Representations

Abstract

Large‐sample deep learning runoff models rely primarily on aggregated catchment attributes, which may obscure informative spatial patterns. Using 531 watersheds in the contiguous United States, we develop a hierarchical deep learning framework that represents static watershed properties from watershed‐averaged raster aggregation to convolutional neural network encodings at fixed‐grid and adaptive multi‐patch scales, thereby providing a controlled test of whether increasingly explicit spatial encoding improves temporal generalization and spatial transferability to ungauged watersheds. Watershed‐averaged rasters perform comparably to curated attribute tables and show a slight advantage under the prediction‐in‐ungauged‐basins benchmark. Incorporating intermediate spatial structure via fixed‐grid encoding provides the clearest temporal‐split benefit relative to watershed‐averaged aggregation, with 26.2% of watersheds showing uncertainty‐supported Nash–Sutcliffe efficiency improvement and 6.0% showing uncertainty‐supported degradation, although this advantage weakens under spatial holdout. Under spatial holdout, diagnostic analyses indicate that fixed‐grid encoding tends to outperform watershed‐averaged raster aggregation in higher‐elevation, steeper watersheds. The adaptive multi‐patch level further reveals that finer spatial representation does not necessarily provide consistent gains, highlighting a trade‐off between spatial detail and computational demand. Overall, this framework provides practical guidance for selecting raster‐based spatial representations in large‐sample deep learning runoff prediction.