Unified multi-task learning for hydrological processes using a shared transformer framework.

Area of Science:

• Hydrology
• Deep Learning
• Environmental Science

Background:

• Most deep learning studies in hydrology use single-task frameworks, limiting cross-process learning.
• Independent modeling of variables like rainfall and streamflow misses synergistic opportunities.

Purpose of the Study:

• Introduce a unified multi-task, multi-modal deep learning framework for hydrology.
• Enable joint streamflow forecasting and rainfall temporal super-resolution within a shared architecture.
• Compare multi-task learning performance against single-task configurations.

Main Methods:

• Developed a shared Transformer encoder with task-specific decoders.
• Integrated temporal and spatial hydrological data.
• Trained a unified model for streamflow forecasting and rainfall super-resolution.
• Compared joint optimization against individually trained models for each task.

Main Results:

• Multi-task training significantly improved streamflow forecasting accuracy.
• Rainfall reconstruction performance remained comparable to single-task models.
• Demonstrated stable streamflow forecasts and hydrologically consistent rainfall reconstructions.

Conclusions:

• Unified, process-aware architectures can effectively represent multiple hydrological components.
• Multi-task learning offers a promising approach for enhancing hydrological modeling.
• The framework shows potential for integrated analysis of the hydrological cycle.