Downscaling of AMSR-E soil moisture with MODIS products using machine learning approaches

Passive microwave remotely sensed soil moisture products, such as Advanced Microwave Scanning Radiometer on the Earth Observing System (AMSR-E) data, have been routinely used to monitor global soil moisture patterns. However, they are often limited in their ability to provide reliable spatial distribution data for soil moisture due to their coarse spatial resolutions. In this study, three machine learning approaches—random forest, boosted regression trees, and Cubist—were examined for the downscaling of AMSR-E soil moisture (25 × 25 km) data over two regions (South Korea and Australia) with different climatic characteristics using moderate resolution imaging spectroradiometer products (1 km), including surface albedo, land surface temperature (LST), Normalized Difference Vegetation Index, Enhanced Vegetation Index, Leaf Area Index, and evapotranspiration (ET). Results showed that the random forest approach was superior to the other machine learning models for downscaling AMSR-E soil moisture data in terms of the correlation coefficient [r = 0.71/0.84 (South Korea/Australia) for random forest, 0.75/0.77 for boosted regression trees, and 0.70/0.61 for Cubist] and root-mean-square error (RMSE = 0.049/0.057, 0.052/0.078, and 0.051/0.063, respectively) through cross-validation. The ET and LST were identified as the most influential among the six input parameters when estimating AMSR-E soil moisture for South Korea, while ET, albedo, and LST were very useful for Australia. In overall, the downscaled soil moisture with 1 km resolution yielded a higher correlation with in situ observations than the original AMSR-E soil moisture data. The latter appeared higher than the downscaled data in forested areas, possibly due to the overestimation of soil moisture by passive microwave sensors over forests, which implies that downscaling can mitigate such overestimation of soil moisture.