A space-time observation system for soil moisture in agricultural landscapes

Abstract

A confluence of scientific and technological developments in geospatial data have made it possible to parameterise the components of the soil water balance equation in space and time at spatial and temporal resolutions useful for agriculture. In this work, we present results on the development of an approach that takes advantage of this opportunity to predict soil moisture at a spatial resolution of 90 m on a daily time step at multiple depths in the profile.Three types of water balance model were examined: (i) single layer model with saturated flow (ii) multi-layer model with saturated flow and (iii) multi-layer model with unsaturated flow. Five layers were considered: 0–5, 5–15, 15–30, 30–60, and 60–100 cm, which coincide with the layers of the Soil Landscape Grid of Australia which is available at ~90 m spatial resolution. Pedotransfer functions were used to predict the bucket size for each soil layer. Precipitation and evapotranspiration are estimated by gridded SILO precipitation data (5 km, 1 day) and the MODIS 16 ET product (1 km, 8 day), respectively. Soil moisture predictions were tested against four publicly and privately owned soil moisture networks.The multi-layer model incorporating unsaturated flow performed the best in terms of predicting soil moisture for the whole profile (0–1 m) with a median correlation coefficient (r) of just over 0.7 across all sites. When classifying the sites according to the land use; cropping sites showed better median correlation (~0.8) than grazing sites (~0.7). However, grazing sites seem to have more consistent results for all the layers. To understand the relative importance of the water balance model predictions as compared to other environmental properties, a Random Forest model was fitted to a suite of variables e.g. soil order, month, temperature and etc., that vary in space, time or both space and time. For the analysis, only calibrated soil moisture network sites (OzNet) were considered. Soil moisture which was derived from unsaturated soil water balance model was the 9th most important variable. To assess the quality of the predictive model, leave-one-out-site cross validation (LOOSCV) was performed and across all sites the prediction quality was reasonable (Concordance = 0.66, Accuracy = 0.06 cm 3 cm−3). These results highlight the potential of this approach and since it is based on readily available data it is scalable to large spatial domains.