Estimating root-zone soil moisture in Mediterranean vineyards using machine learning

Root-zone soil moisture (RZSM) reflects the water accessible to plants and is therefore central to precision irrigation support and agricultural drought monitoring, yet direct RZSM observations are limited. Satellite missions provide surface soil moisture (SSM), but they do not directly observe deeper layers, and in-situ measurements remain too sparse for broad coverage. We present a machine learning approach to estimate daily RZSM in vineyards in the Terra Alta region of Catalonia in northeastern Spain, using daily 2020 to 2024 in-situ observations from eight stations as reference data. This model provides a baseline for later experiments using satellite SSM to extend applicability beyond the instrumented network.
We train a multilayer perceptron (MLP) to predict soil moisture at 25 cm, taken as RZSM, using in-situ SSM at 5 cm, daily precipitation, mean, minimum and maximum temperature, a cyclic encoding of day of year, and static soil descriptors from SoilGrids. Robustness is assessed with year-block cross-validation to evaluate temporal generalisation and leave-station-out experiments to evaluate transferability across vineyards. Performance is quantified using non-parametric Kling–Gupta efficiency (KGE) and RMSE.
The model achieves strong skill when evaluated on independent years at training stations, with median KGE around 0.9. Transfer to unseen vineyards is more heterogeneous, with some stations retaining good performance around 0.85 and others showing biases and reduced efficiency, suggesting that additional information may be needed for consistent transfer across vineyards. Ongoing work aims to improve generalisation by incorporating antecedent moisture and precipitation information and by testing additional predictors such as vegetation, supported by feature importance analysis across the full set of inputs. To enable use beyond the instrumented network, we will transition the model towards configurations driven by or trained with satellite-derived SSM. Taken together, these steps are intended to move towards a transferable tool to support drought monitoring and irrigation-related decisions in agricultural regions.