Knowledge of soil moisture is crucial in every agricultural system. Accurate soil moisture estimates are valuable guides to management decisions such as crop selection and estimation of yield potential. However, soil moisture is highly variable temporally and spatially in the landscape and very challenging to monitor. Existing soil moisture monitoring approaches lack in appropriate resolution for current agricultural requirements. Therefore, this thesis develops novel space-time models for predicting and forecasting soil moisture at a 90 m spatial resolution.
A water balance model (unsaturated and multi-layer) was presented to estimate soil moisture with a combination of soil moisture sensing, remote sensing and readily available geospatial data, which parameterise the soil water balance equation. The model was further improved by optimising the parameters (infiltration and evapotranspiration) and using machine learning techniques. The prediction quality was reasonable: topsoil (Concordance = 0.69, Accuracy = 0.05 cm3cm-3); subsoil (Concordance = 0.72, Accuracy = 0.04 cm3cm-3); and root-zone (Concordance = 0.75, Accuracy = 0.05 cm3cm-3).
Also, three farm-scale soil moisture surveys were performed using a mobile cosmic-ray probe platform at strategic time points of the farming year. Estimates were validated with the soil moisture obtained from 0-30 cm soil layer (Concordance = 0.87, Accuracy = 0.05 cm3cm-3). Moreover, a model (stochastic plus machine learning) was presented to forecast soil moisture: one month; three months; and six months into the future. One month lead times had the greatest forecast quality (Concordance > 0.9, Accuracy < 0.02) with the forecast quality declining with longer lead times. The subsoil was more forecastable than the topsoil. Thus, these new approaches for soil moisture modelling would have real utility across agricultural regions of Australia where they are likely to be most beneficial in management of agri-food production systems.