Multi-source machine learning and spaceborne remote sensing data accurately predict three-dimensional soil moisture...

Area of Science:

• Environmental Science
• Remote Sensing
• Geospatial Analysis

Background:

• Arid and semi-arid environments are used for waste storage due to low groundwater recharge, minimizing contaminant transport.
• Evapotranspiration (ET) covers utilize vegetation to reduce water percolation into waste disposal cells.
• Accurate soil moisture (SM) monitoring is crucial for ET cover performance, balancing drainage and radon flux.

Purpose of the Study:

• To investigate the potential of non-invasive radar remote sensing and geospatial data for monitoring soil moisture (SM) in ET covers.
• To develop and evaluate machine learning (ML) models for estimating the SM profile in vegetated disposal cells.
• To provide a practical, accurate, and spatially comprehensive tool for SM monitoring.

Main Methods:

• Theoretical simulations analyzed multi-frequency radar backscatter sensitivity to SM at various depths.
• Shallow and deep machine learning (ML) models were developed using Google Earth Engine to estimate SM profiles.
• Models integrated satellite radar, optical/infrared sensor data, and rainfall data, trained and validated with in situ SM measurements.

Main Results:

• Lower frequency radar (L- and P-band) showed better sensitivity to deeper soil layers and a larger SM dynamic range.
• ML models accurately estimated SM across six soil layers (0-2 m), with high correlation (r=0.75–0.94) and low error.
• A simpler shallow-learning ML approach outperformed a deep-learning model in accuracy.

Conclusions:

• Non-invasive radar remote sensing combined with ML provides an accurate and cost-effective method for monitoring SM in ET covers.
• This approach offers a spatially comprehensive alternative to traditional invasive instrumentation.
• The developed ML models can be applied to other vegetated and potentially rock-armored disposal cell covers.