Continental scale downscaling of AWRA-L analysed soil moisture using random forest regression

Abstract

The Australian Water Resource Assessment Landscape (AWRA-L) model as used by the Bureau of Meteorology (BoM) provides daily continental scale soil moisture (SM) estimates (among other landscape water variables) at ~5-km resolution. At such a coarse scale these data cannot represent the high spatiotemporal variability of SM across heterogeneous land surfaces. Downscaling of coarse SM products based on machine learning (ML) has become increasingly popular due to its robust predictions and potential for large-scale applications. As a first step towards high-resolution daily Australia-wide SM estimation, a downscaling framework was developed to generate monthly SM with 500-m spatial resolution using analysed SM from AWRA-L and multisource geospatial predictors in random forest (RF) regression. Candidate predictors include digital elevation model (DEM), soil properties from the Australian soil and landscape grids, and several retrievals from the MODerate-resolution Imaging Spectroradiometer (MODIS). Ten experiments were conducted to decide the best combination of predictors. In the chosen model, DEM and available water capacity (AWC) were consistently identified as the most important predictors based on the ranking of variable importance. The downscaled SM shows greatly enhanced spatial details at the local scale while maintaining consistent patterns with AWRA-L analysis at the continental scale. Validations against in-situ measurement networks using Pearson correlation coefficient (R) show that there is very little difference in the performance between the downscaled and AWRA-L SM. Average R values for the downscaled SM against CosmOz, OzFlux and OzNet were 0.87, 0.68 and 0.75, respectively, while the original AWRA-L SM average R were 0.86, 0.68 and 0.76, respectively. Furthermore, the time series comparison based on a wetness unit shows that the downscaled SM can well catch up the fluctuations of in-situ SM. In general, this study explores the potential of ML approach for the SM downscaling applications at the continental scale. It could be a promising direction to exploit the modelling capability of integrating multisource geospatial data including satellite retrievals, land surface models (LSM) and interpolated ground observation data. Future directions should concentrate on integrating this approach into an operational framework with a daily frequency. Exploration of the relationships between SM and auxiliaries under difference scales would be essential, in order to better understand the dominant physical controls on spatial variability of SM. Show more