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

Area of Science:

• Hydrology
• Environmental Science
• Data Science

Background:

• Accurate multi-step runoff forecasting is crucial for China's flood control, water resource management, and hydrological assessments.
• Existing data-driven methods face challenges in capturing diverse temporal scales and river network dependencies, limiting accuracy and spatial consistency.

Purpose of the Study:

• To develop an advanced spatiotemporal forecasting framework to enhance runoff prediction accuracy and spatial structural consistency.
• To address the limitations of current methods in integrating multi-scale temporal variations and river network topology.

Main Methods:

• A novel spatiotemporal forecasting framework combining cross-scale temporal fusion and basin-topology-guided spatial modeling.
• Implementation of a multi-scale temporal module with cross-scale gating for adaptive integration of runoff variations.
• Utilization of a basin-topology attention module to incorporate upstream-downstream connectivity in spatial dependency learning.

Main Results:

• The proposed method demonstrated superior performance over existing baselines on a China-scale gridded runoff forecasting benchmark.
• Achieved improved metrics including MAE (0.0269) and RMSE (0.0603) in normalized log-scale runoff units, with PSNR 24.39 and SSIM 0.9273.
• The framework effectively reduced numerical errors and preserved spatial patterns of runoff fields with practical inference efficiency.

Conclusions:

• The developed spatiotemporal framework significantly enhances runoff forecasting accuracy and spatial consistency in China.
• Both cross-scale gating and basin-topology attention mechanisms were confirmed to be vital contributors to the overall performance improvement.
• The findings support the framework's utility for improved hydrological assessments and water resource management.