Improving watershed-scale daily nutrient simulation using a process-model-informed graph attention network with...

Area of Science:

• Environmental Science
• Hydrology
• Artificial Intelligence

Background:

• High-frequency water quality monitoring is crucial for effective watershed management.
• Sparse monitoring stations and spatial heterogeneity in watersheds pose significant challenges.
• Existing methods struggle to provide accurate predictions at ungauged locations.

Purpose of the Study:

• To develop an advanced AI model for improved water quality prediction in watersheds.
• To integrate in situ observations, process-based model outputs, and satellite data for enhanced accuracy.
• To improve water quality simulations at ungauged river reaches through data assimilation and transfer learning.

Main Methods:

• A Process-Model-Informed Graph Attention Network (PMIGAT) was developed.
• Intermittent satellite-retrieved water quality data were used as weak supervision.
• A similarity-guided graph attention module facilitated targeted information transfer from monitored to ungauged reaches.

Main Results:

• PMIGAT achieved KGE of 0.66 at monitored reaches and 0.60 at sparsely gauged reaches with satellite data.
• On ungauged reaches without satellite data, PMIGAT significantly improved R² (0.01 to 0.46) and reduced MAPE (64% to 26%) compared to SWAT.
• High-concentration event detection improved, with CSI increasing from 0.04 to 0.28 and relative peak error decreasing from 60% to 13%.

Conclusions:

• The PMIGAT model effectively generates daily water quality data, even with intermittent monitoring.
• Satellite data and the similarity-guided graph attention module synergistically enhance predictions at sparsely monitored and ungauged locations.
• The method supports accurate hotspot identification and robust watershed management strategies.