From Soil Moisture Patterns to Hydrological Connectivity: An Explainable AI Approach for Nitrate Modeling

Hydrological connectivity is crucial for the mobilization, transport, and transformation of nitrate, but quantifying it at the catchment scale remains challenging, especially when capturing the spatial features that influence hydrological transport. We address this challenge by leveraging SMAP-Hydroblocks (Vergopolan et al., 2021), a high-resolution soil moisture dataset, to explore spatial soil moisture patterns as proxies for hydrological connectivity by predicting stream nitrate concentrations. We simulated daily nitrate concentrations across nine U.S. catchments with diverse land cover and concentration-discharge (C-Q) relationships using a multi-branch deep learning. We trained the model on discharge time series as an aggregated measure of hydrological connectivity, soil moisture spatial patterns to account spatial heterogeneities that influence hydrological connectivity, height above the nearest network maps as spatial flopath indicator and static proxies of nitrogen sources (nitrogen surplus and fraction of urban areas of catchments). 

Our model achieved robust performance, with a median Nash-Sutcliffe Efficiency (NSE) of 0.63 and a median Kling-Gupta Efficiency (KGE) of 0.74 across the test period, outperforming traditional C-Q relationship models. Explainable AI (XAI) techniques revealed that spatial patterns of soil moisture contribute significantly to nitrate predictions, accounting for 30% of feature importance on average. Excluding these patterns decreased model accuracy by 14%. Explainable AI (XAI) methods revealed distinct hydrological responses across catchments: in catchments with positive C-Q patterns, spatial soil moisture patterns amplified nitrate transport during wet periods, while discharge dilution effects are more important in catchments with negative C-Q relationships. Attention maps highlighted near-stream zones as critical areas for predicting nitrate transport, reflecting their dominant role in hydrological connectivity and nitrate dynamics.

This study demonstrates the potential of integrating deep learning, XAI, and remote sensing products to quantify hydrological connectivity and nitrate dynamics. These findings provide new insights into the spatial and temporal variability of nitrate transport across catchments and a framework for improving water quality management.

Vergopolan, N., Chaney, N. W., Pan, M., Sheffield, J., Beck, H. E., Ferguson, C. R., Torres-Rojas, L., Sadri, S., & Wood, E. F. (2021). SMAP-HydroBlocks, a 30-m satellite-based soil moisture dataset for the conterminous US. Scientific Data, 8(1), 1. https://doi.org/10.1038/s41597-021-01050-2