Estimating volumetric water content from electrical resistivity using a random forest model

TL;DR

This study develops random forest models to estimate volumetric water content from electrical resistivity tomography data, improving accuracy with environmental parameters but highlighting challenges in model generalization across locations.

Contribution

The paper introduces a novel application of random forest models incorporating precipitation and temperature to predict VWC from ERT data, addressing heterogeneity and temporal dynamics.

Findings

Including precipitation and temperature reduced MAE significantly.

Depth-dependent delays improved maximum and mean relative errors.

Models trained on one location did not generalize well to others.

Abstract

Background: Accurately estimating volumetric water content (VWC) can greatly enhance the prediction of landslide risk. The standard approach involves using locally limited, invasive sensor measurements. Recently, however, electrical resistivity tomography (ERT) has emerged as a cost-effective, minimally invasive, real-time, indirect method of monitoring VWC. However, linking ER to VWC poses distinct challenges. Purpose: Random forest models were developed to estimate and predict VWC from ER. Spatially and temporally heterogeneous measurements were conducted to improve robustness and accuracy. Methods: The models were trained using 370-500 ER sensor measurements at depths of 10, 50, 150 and 190 cm. Precipitation and air temperature with a depth-dependent impact delay were introduced to simulate infiltration. The models were then validated at various locations, and their predictive capabilities were examined. Results: Including precipitation and temperature as parameters reduced the mean absolute error (MAE) by between 7.3% and 73.0%, depending on the depth. However, not all outliers could be eliminated. Delaying the model parameter depth-dependently reduced the maximum relative error (MaRE) by 14.7-56.0% and the mean relative error (MRE) to below 2.8%. Observed uncertainties for the top meter remained fairly constant. A 24 h forecast revealed an MRE below 5% for all depths; however, outliers occurred more frequently. Nevertheless, models trained on data from one location failed to generalize to other locations. VWC tomography based on ERT revealed significant moisture bands at depths of one to two meters. Conclusion: VWC can be reliably predicted from ERT. However, validation of VWC from ERT in different locations revealed significant deficits. It is expected that an automated ERT training of models at various locations will eliminate these differences.