China runoff-field forecasting based on cross-scale gating and basin-topology attention.

Abstract

Accurate multi-step runoff forecasting over China is important for flood control, water-resource management, and regional hydrological assessment. However, existing data-driven methods often struggle to jointly capture temporal variations at different time scales and the directional hydrological dependencies imposed by river networks, which limits forecasting accuracy and spatial structural consistency. To address this issue, this paper proposes a spatiotemporal forecasting framework that combines cross-scale temporal fusion with basin-topology-guided spatial modeling. Specifically, a multi-scale temporal module with cross-scale gating is introduced to adaptively integrate short-, medium-, and long-term runoff variations, while a basin-topology attention module incorporates upstream-downstream connectivity into spatial dependency learning. Experiments are conducted on a China-scale gridded runoff forecasting benchmark derived from the publicly available GloFAS Historical dataset through spatial filtering, valid-region masking, and forecasting-oriented sample construction. The proposed method achieves better overall performance than representative baselines in terms of MAE, RMSE, PSNR, and SSIM. In the overall comparison, it reaches MAE 0.0269 and RMSE 0.0603 in normalized log-scale runoff units, PSNR 24.39, and SSIM 0.9273, while maintaining a moderate parameter size and practical inference efficiency. The results demonstrate that the proposed framework reduces numerical errors and better preserves the spatial patterns of runoff fields. Ablation studies further confirm that both cross-scale gating and basin-topology attention contribute consistently to the overall improvement.