Improving 5TE Soil Water Content Sensor Accuracy Using Neutron Probe Calibration Data and Machine Learning

Accurate soil volumetric water content (VWC, m 3 m − 3 ) measurements are crucial for agricultural and environmental applications, yet capacitance-based sensors like 5TE often exhibit systematic errors. This study employs machine learning (ML) to enhance the accuracy of the 5TE VWC readings, using co-located Neutron Probe (NP) VWC readings collected over nine growing seasons ( 2017 − 2025 ) at four stations as reference. Four ML models: ridge regression, Gaussian process regression (GPR), random forest (RF), and eXtreme Gradient Boosting (XGBoost) are compared. Inputs to the models are restricted to quantities the 5TE already reports such as soil temperature and soil electrical conductivity, together with installation depth, day of year and the timing and magnitude of wetting events, requiring no site-specific measurement of soil texture or bulk density. Model generalization was evaluated using three blocked cross-validation strategies—leave-one-year-out (LOYO), leave-one-station-out (LOSO), and their combination leave-one-year-one-station-out (LOYOSO)—with hyperparameter tuning nested within the block. Temporal generalization proved the most demanding condition: LOYO produced the highest RMSE for every model ( 0.066 − 0.074 m 3 m − 3 ), followed by LOSO ( 0.062 − 0.068 ) and LOYOSO ( 0.055 − 0.069 ). On the same held-out folds, all models achieved a lower RMSE than the uncorrected 5TE readings ( 0.084 − 0.099 m 3 m − 3 ). LOSO errors were the most consistent across folds, indicating that transfer to an unseen station under the same weather pattern is comparatively tractable. All models showed negligible bias across all strategies ( | MBE | ≤ 0.006 m 3 m − 3 ). The trained models were applied forward to routine 5TE readings to produce continuous VWC records. Validation is confined to a single site, soil texture and cropping system, and transferability to contrasting conditions remains to be established.