In the context of smart agriculture, accurate soil moisture monitoring is crucial to optimise irrigation, improve water usage efficiency and increase crop yields. Although low-cost capacitive sensors are used to make monitoring affordable, these sensors face accuracy challenges that often result in inefficient irrigation practices. This paper presents a method for calibrating capacitive soil moisture sensors through data assimilation. The method was validated using data collected from a farm in Dos Hermanas, Seville, Spain, which utilises a drip irrigation system. The proposed solution integrates the Hydrus 1D model with particle filter (PF) and the Iterative Ensemble Smoother (IES) to continuously update and refine the model and sensor calibration parameters. The methodology includes the implementation of physical constraints, ensuring that the updated parameters remain within physically plausible ranges. Soil moisture was measured using low-cost SoilWatch 10 capacitive sensors and ThetaProbe ML3 high-precision sensors as a reference. Furthermore, a comparison was carried out between the PF and IES methods. The results demonstrate that the data assimilation approach markedly enhances the precision of sensor readings, aligning them closely with reference measurements and model simulations. The PF method demonstrated superior performance, achieving an 84.8% improvement in accuracy compared to the raw sensor readings. This substantial improvement was measured against high-precision reference sensors, confirming the effectiveness of the PF method in calibrating low-cost capacitive sensors. In contrast, the IES method showed a 68% improvement in accuracy, which, while still considerable, was outperformed by the PF. By effectively mitigating observation noise and sensor biases, this approach proves robust and practical for large-scale implementations in precision agriculture.

This article reports the field validation of an Adaptive Model Predictive Controller (AMPC) for precision irrigation on a commercial chard crop in Dos Hermanas (Seville, Spain). The system operates on a long range wide area network (LoRaWAN) based architecture with dual-depth soil moisture low-cost nodes sensors and remote actuation system composed of a pump and on/off valves. The AMPC employs a discrete linear time-varying model, whose parameters will be updated online using Recursive Least Squares with a forgetting factor. At each sampling instant, the controller solves a constrained optimisation problem with soft constraints for the moisture boundaries and penalties on irrigation pulse changes and pump switching. Over a 24-day experimental period, the AMPC kept soil moisture predominantly within agronomically prescribed bounds, with brief excursions attributable to rainfall events beyond the controller's authority. Relative to a manually irrigated plot of the same size, it reduced water use with the same number of pump starts. This case-study deployment demonstrates the feasibility and robust operation of AMPC under real farm conditions, supporting the potential adoption of AMPC in commercial irrigation systems. Scope. This is a single-site case study; results are case-specific and not statistically generalisable. We therefore restrict claims to the directly validated outcomes (soil-moisture regulation and irrigation-water reduction) and we mention agronomic impacts (e.g., yield) as observational context only.

This paper presents a multi-layer modelling methodology for simulating water dynamics in unsaturated soils, designed to maintain a conservative balance and reduce computational costs compared to continuous models based on Richards’ equation. The model incorporates infiltration, evaporation, capillary redistribution, transpiration and gravitational drainage using an explicit time integration scheme. Technical validation is carried out by comparison with HYDRUS-1D, simulating a complete 120-day strawberry growing cycle under identical forcing conditions and hydraulic parameters. The results show high dynamic agreement, achieving an RMSE=1.19mm and a reduction in computation time of over 97%. Subsequently, the methodology was experimentally evaluated in April 2026 at a greenhouse strawberry plantation in Moguer (Huelva), using actual soil moisture measurements at different depths. The results obtained demonstrate the model’s good predictive capability, achieving an RMSE=4.71%VWC and a Willmott index dW=0.833 for the cumulative profile.