Bodemvocht schatten met SWIM²: Inverse modellering met bodemvochtsensoren voor verbeterde voorspellingen van irrigatiebehoefte

Improved irrigation management is a key strategy to increase agricultural water use efficiency, particularly in the face of increasing drought risk. Soil moisture is a critical variable for decision-making in irrigation scheduling, yet its measurement and prediction remain uncertain, especially when relying on sparse soil sensors at fixed locations. This dissertation aims to enhance real-time soil moisture monitoring and forecasting for irrigation support by quantifying sensor uncertainty, integrating field measurements into a probabilistic modeling framework, and developing a digital twin as decision support system (DSS). The first part of this dissertation focuses on the fundamental role of observational error correlation in sensor measurements and its impact on parameter estimation for soil moisture modeling. First, we demonstrate how temporal autocorrelation of observational errors can significantly affect parameter uncertainties through both analytical and Bayesian approaches (Generalized Least Squares and DREAM(ZS)). Neglecting this error autocorrelation may lead to either over- or underestimation of predictive uncertainty. Next, we explore the physical origin of this temporal autocorrelation in observational errors. We propose a mechanistic error model grounded in water retention variability, assuming a horizontally uniform soil water potential, allowing the prediction of spatiotemporal variance in soil moisture based on physical soil properties. Field and numerical experiments confirm that soil water retention variability can explain persistent, autocorrelated deviations in sensor readings from the spatial average, with implications for parameter inference and uncertainty estimation. Since we do not always have the means to measure or know variability in physical soil properties, and sensor-based soil moisture error structures cannot always be characterized per field, we introduce a generalized error modeling framework. Using data from 93 cropping cycles across three growing seasons in Belgium, we quantify pooled systematic and random errors in soil moisture content estimates from TEROS 10 sensors. Composite soil sampling provided unbiased reference estimates of field-scale moisture content. From these, we derived a pooled sensor calibration function to represent the top 30 cm soil layer, and estimated the temporal correlation of the observational errors. The resulting pooled error covariance matrix showed strong autocorrelation (0.52), driven by systematic sensor offsets, and challenges the common assumption of uncorrelated sensor errors in hydrological modeling. The second part of this dissertation translates these insights into a practical framework for real-time soil moisture forecasting and irrigation support. We present SWIM² (Sensor Wielded Inverse Modeling of a Soil Water Irrigation Model), a digital twin integrating biased sensor data, unbiased soil samples, and meteorological inputs into a Bayesian inverse modeling framework. SWIM² estimates soil and crop parameters and provides real-time probabilistic forecasts of soil moisture with quantified uncertainty. Field validation on 18 cropping cycles showed that SWIM² delivers robust 7-day forecasts, with 4-day predictive accuracy similar to that of the sensors themselves. The importance of soil sampling and error correlation knowledge was reaffirmed: samples correct for sensor offset, while sensor data are essential for tracking dynamic changes, yet neglecting error correlation can lead to biased parameter estimates and overconfident, misleading forecasts. Finally, we integrate weather forecasts into SWIM², assessing the impact of both deterministic and ensemble-based weather forecasts, to support a real-time irrigation DSS. SWIM² predicts irrigation needs by forecasting soil moisture relative to crop-specific thresholds. Both weather forecast types led to high accuracy and recall in predicting irrigation triggers, with ensemble forecasts offering greater benefit at longer lead times. Model simulations of irrigation treatments further demonstrated the DSS's potential to optimize water use efficiency and assess crop response under different management strategies. In summary, this dissertation presents a comprehensive framework for uncertainty-aware soil moisture monitoring and modeling, and real-time irrigation decision support. It quantifies sensor uncertainty, integrates these uncertainties into a Bayesian modeling workflow, and demonstrates a scalable solution for site-specific irrigation scheduling. These contributions lay the groundwork for more resilient and efficient water use in agriculture, guided by digital tools and sensor-driven data science.