Trade-offs between data-driven and process-based approaches for root-zone soil moisture retrieval in a Mediterranean vineyard

Accurate monitoring of water availability in the root zone is a prerequisite for generating precise irrigation recommendations and mitigating drought impacts in water-limited Mediterranean ecosystems. This work evaluates the performance and physical consistency of two distinct modelling paradigms to retrieve Root-Zone Soil Moisture (RZSM) in a vineyard located in Terra Alta (Catalonia, Spain), intended as a basis for operational decision support.

We contrast a purely data-driven method, utilizing a Multilayer Perceptron (MLP), against a process-based approach that couples a parsimonious multilayer soil model with an Ensemble Kalman Filter (EnKF) for the assimilation of Surface Soil Moisture (SSM). Both schemes are currently benchmarked using in-situ SSM observations and standard meteorological forcing.

The results highlight a clear dichotomy between predictive skill and physical interpretability. The neural network approach demonstrated excellent performance in capturing non-linear seasonal trends and rapid wetting events, yielding better Kling–Gupta Efficiency (KGE) scores during validation. Conversely, the physical model exhibited lower statistical metrics but ensured mass conservation and provided a transparent representation of vertical water transport.

We conclude that while machine learning excels in reproducing local dynamics, the physical framework offers the robustness required for consistent water accounting. Consequently, we propose a synergistic roadmap where machine learning is leveraged to regionalize model parameters, and the assimilation of high-resolution satellite Surface Soil Moisture serves to spatialize the state estimates. This integration is essential to scale up from plot-level findings to regional irrigation recommendations, supporting the next generation of Digital Twins in agriculture.