Saavedra et al. (2025) From Soil Moisture Spatial Patterns to Catchment Nitrate Dynamics Using Explainable AI

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2025
• Date: 2025-11-01
• Authors: Felipe Saavedra, Noemi Vergopolan, Andréas Musolff, Ralf Merz, Zhenyu Wang, Carolin Winter, Larisa Tarasova
• DOI: 10.1029/2025wr040295

Research Groups

Not specified in the provided abstract.

Short Summary

This study developed a multi-branch Deep Learning framework, leveraging high-resolution satellite soil moisture data, to predict daily nitrate concentrations in streams across eight US catchments. The model successfully represents nitrate dynamics, demonstrating that spatial patterns of soil moisture are significant predictors and identifying near-stream hotspots as critical areas for nitrate export.

Objective

• To develop a multi-branch Deep Learning framework to predict daily nitrate concentrations in streams at the catchment scale.
• To test the hypothesis that spatial patterns of soil moisture are indicative of hydrological connectivity and can be used to predict nitrate concentrations in streams.

Study Configuration

• Spatial Scale: Eight catchments across the US; soil moisture data at 30 m resolution, aggregated to 1 km.
• Temporal Scale: Daily.

Methodology and Data

• Models used: Multi-branch Deep Learning framework, Explainable AI (XAI).
• Data sources: Streamflow observations, SMAP-Hydroblocks (satellite-based soil moisture data at 30 m resolution, aggregated to 1 km).

Main Results

• The developed model satisfactorily represents nitrate dynamics in the study catchments, achieving a median Nash-Sutcliffe Efficiency of 0.62.
• Spatial patterns of soil moisture account for an average of 32% of the model's feature importance, and model performance deteriorates by an average of 14% if these patterns are excluded.
• Explainable AI confirmed that model decisions align with known physical processes across catchments with contrasting concentration-discharge (C-Q) behavior.
• Attention maps identified near-stream hotspots as regions with the highest predictive power for nitrate export, suggesting a link between soil moisture patterns and catchment-scale hydrological connectivity.

Contributions

• Development of a novel multi-branch Deep Learning framework for predicting daily nitrate concentrations at the catchment scale.
• Demonstration of the significant role of high-resolution satellite-based soil moisture spatial patterns in improving nitrate predictions.
• Application of Explainable AI to confirm the physical consistency of the model and identify critical near-stream hotspots for nitrate export.
• Proof-of-concept for combining XAI with high-resolution remote sensing products to enhance nitrate predictions and map critical areas for export.

Funding

Not specified in the provided abstract.

Citation

@article{Saavedra2025From,
  author = {Saavedra, Felipe and Vergopolan, Noemi and Musolff, Andréas and Merz, Ralf and Wang, Zhenyu and Winter, Carolin and Tarasova, Larisa},
  title = {From Soil Moisture Spatial Patterns to Catchment Nitrate Dynamics Using Explainable AI},
  journal = {Water Resources Research},
  year = {2025},
  doi = {10.1029/2025wr040295},
  url = {https://doi.org/10.1029/2025wr040295}
}