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

Water-quality management requires high-frequency monitoring data, which remains challenging, especially in watersheds exhibit pronounced spatial heterogeneity and sparse monitoring stations. To address this gap, this study proposes a Process-Model-Informed Graph Attention Network (PMIGAT) that integrates in situ observations and process-based variables from a process-based model, and implements intermittent satellite-retrieved water quality data as weak supervision to improve predictions at ungauged reaches. A similarity-guided graph attention module is further introduced to enable targeted transfer of supervisory information from monitored nodes to ungauged reaches based on hydrological and landscape similarity. The proposed method was evaluated for nitrogen simulation in the Hangbu River Basin, China. Results showed that the Kling-Gupta efficiency (KGE) at continuously monitored reaches was 0.66, and the median KGE at sparsely gauged reaches reached 0.60 on dates with satellite retrievals. On ungauged reaches on dates without satellite retrievals, PMIGAT outperformed the process-based model, such as Soil and Water Assessment Tool (SWAT), increasing R² from 0.01 to 0.46 and reducing the mean absolute percentage error (MAPE) from 64% to 26%. Furthermore, the new method also improved the detection of high-concentration events with critical success index increasing from 0.04 to 0.28, and the relative peak error decreasing from 60% to 13%. Ablation analyses indicated that satellite retrievals contributed the largest gains at sparsely gauged reaches, and its synergy with similarity-guided graph attention module strengthened with higher satellite availability, shorter along-river distance to the outlet, and greater land-surface similarity. The method can generate spatiotemporally daily water-quality data despite intermittent monitoring, supporting accurate hotspot identification and watershed management.