Multi-depth soil moisture dynamics to rainfall events: An automated machine learning approach

Abstract. This study proposes an integrated, event-based framework for quantifying soil moisture dynamics at multiple depths (10, 20, 30, and 40 cm) in response to rainfall events using an automated machine learning (AutoML) approach. At the observatory we record the hydrometeorological and soil moisture data at different depth below the ground surface at every 10-minute intervals. We use these datasets to capture both rapid single-peak and gradual multiple-peak soil moisture responses during diverse rainfall events.

Recognising that manual model selection and hyperparameter tuning are labour intensive and may not fully capture the complex, non-linear interactions among hydrometeorological variables. Here we propose an AutoML framework that leverages Bayesian optimisation to predict subsurface soil moisture at different depths. The model was evaluated under four temporal scenarios: S1 (March–May), S2 (March–June), S3 (March–July), and S4 (March–August), for the full dataset and rainfall-only instances, separately. This automatic selection and tuning of various regression models result in superior predictive performance as compared to benchmark algorithms. The coefficients of determination ranges from 0.88 to 0.98 and minimal root mean squared errors (1.6 %–3.4 %). Further, the global sensitivity analysis indicates that the atmospheric humidity and dew point strongly influence near-surface moisture. The solar radiance and evaporation drive moisture depletion, and soil temperature gradients play a critical role in the vertical profile of the soil column. These findings highlight the value of integrating advanced AutoML techniques with event‐based hydrological analysis to enhance our understanding of soil moisture variability, which has significant implications for water resource management, agricultural planning, and hazard mitigation in variable climatic regimes.