Data-driven modeling of soil water content using different rainfall predictors

Abstract

Soil moisture is a crucial component of hydrological and geotechnical processes, especially for shallow slope failures. When adopting a physically-based modeling approachfor the prediction of hydrological variables, it can however be challenging to gather the necessary data and findwhat parameters to use on a large scale. The goal of this study is therefore to adopt a data-riven approach to predict the amount of water in the soil using only rainfall based predictors. The study used a Random Forest algorithm that was trained using only cumulative rainfall indices and time markers. The analysis was performed independently for the two distinct soil management practices (Control and Rolling) at depths of 10, 50, and 90cm. A randomized validation test was done and the results show that the prediction of water content has very high internal consistency (R2 > 0.98). However, the predictions for the year 2024, which had not been used for the trainingof the model, showed clear physical limits.The reason of this poor perfomance is related to meterological conditions on January-April 2024 period, which differed from the ones present in training time series. Consequently, the results highlight the importance of reconstructing data-driven models with a training time- series long enough to consider all the possible meterological conditions. Nevertheless, the results show that the model can find important near-saturation conditions with a small margin of error.This means that this Random Forest model can be reliable in predicting conditions that could lead to shallow landslide triggering, even while neglecting air temperature data.