Estimating catchment scale soil moisture at a high spatial resolution: Integrating remote sensing and machine learning

Highlights: • High spatial resolution soil moisture is important for a number of applications. • A regression tree, an ANN and a GPR model were developed to downscale soil moisture. • Downscaling models were developed based on the soil thermal inertia theory. • Coarse spatial resolution soil moisture was downscaled to 1 km. • The results, especially from regression tree and GPR models, are very encouraging. Soil moisture information is important for a wide range of applications including hydrologic modelling, climatic modelling and agriculture. L-band passive microwave satellite remote sensing is the most feasible option to estimate near-surface soil moisture (~0–5 cm soil depth) over large extents, but its coarse resolution (~10s of km) means that it is unable to capture the variability of soil moisture in detail. Therefore, different downscaling methods have been tested as a solution to meet the demand for high spatial resolution soil moisture. Downscaling algorithms based on the soil thermal inertia relationship between diurnal soil temperature difference (ΔT) and daily mean soil moisture content (μ_SM) have shown promising results over arid and semi-arid landscapes. However, the linearity of these algorithms is affected by factors such as vegetation, soil texture and meteorology in a complex manner. This study tested a (i) Regression Tree (RT), an Artificial Neural Network (ANN), and a Gaussian Process Regression (GPR) model based on the soil thermal inertia theory over a semi-arid agricultural landscape in Australia, given the ability of machine learning algorithms to capture complex, non-linear relationships between predictors and responses. Downscaled soil moisture from the RT, ANN and GPR models showed root mean square errors (RMSEs) of 0.03, 0.09 and 0.07 cm³/cm³ compared to airborne retrievals and unbiased RMSEs (ubRMSEs) of 0.07, 0.08 and 0.05 cm³/cm³ compared to in-situ observations, respectively. The study showed encouraging results to integrate machine learning techniques in estimating near-surface soil moisture at a high spatial resolution.