From data to interpretable models: machine learning for soil moisture forecasting

Area of Science:

• Hydrology and Soil Science
• Data-driven modeling
• Environmental monitoring

Background:

• Soil moisture is vital for agriculture, ecosystems, and natural disaster prediction.
• Existing soil moisture models have limitations in forecasting accuracy and time horizons.
• Instrumental noise and changing environmental conditions complicate soil moisture monitoring.

Purpose of the Study:

• To introduce and evaluate two novel data-driven soil moisture models: NAR and AEAR.
• To improve the accuracy and extend the forecasting time horizon for soil moisture predictions.
• To investigate the physical interpretability of model parameters in unsaturated hydrological processes.

Main Methods:

• Development of Naive Accumulative Representation (NAR) and Additive Exponential Accumulative Representation (AEAR) models.
• Validation using field data from a post-wildfire site in southern California with varying soil textures and depths.
• Comparison with established and state-of-the-art baseline models in forecasting experiments.
• Controlled laboratory experiments to assess model robustness.

Main Results:

• NAR and AEAR models demonstrate competitive performance against existing forecasting methods.
• The AEAR model accurately fits soil moisture data across different soil textures and depths (5, 15, 30 cm).
• AEAR provides more accurate soil moisture forecasts for time horizons of 10-24 hours compared to existing models.

Conclusions:

• The AEAR model offers a significant advancement in soil moisture forecasting, particularly for extended periods.
• Learned parameters in AEAR represent physically meaningful hydrological processes (gravity, suction).
• Improved long-range soil moisture forecasts from AEAR can provide actionable insights for natural disaster preparedness, reducing potential loss of life and property.