Downscaling SMAP Soil Moisture Estimation with Gradient Boosting Decision Tree Regression over the Tibetan Plateau


The Soil Moisture Active Passive (SMAP) satellite can no longer directly deliver high-resolution (9 km) soil moisture products with the failure of the onboard L-band radar. Thus, an appropriate replacement sensor and new algorithms are urgently needed to compensate for the lost radar and estimate soil moisture at high-resolution. This paper presents a new downscaling approach, the Downscaling basEd oN gradient boosting deciSion trEe (DENSE) method to solve this problem. We analyzed 26 soil moisture related indices, derived from MODIS and a digital elevation model, to identify proxy variables for soil moisture variability. A gradient boosting decision tree regression links the aggregated soil moisture proxies and SMAP observations at a coarse scale to express the nonlinear relationships between them. High-resolution soil moisture products were generated by applying this built regression model to the optimal soil moisture proxies at a fine scale over the entire Tibetan Plateau during the years 2015– 2017. In situ measurements were collected from the Ngari, Naqu, and Maqu networks, which represent different climatic and vegetation conditions. We evaluated the downscaled soil moisture against ground observations at the daily, point, and network scales. The results indicated that the DENSE method effectively infers the spatio-temporal variability of soil moisture; and at the same time, preserves SMAP soil moisture accuracy. The performance of the proposed method failed in the Maqu network as this area is covered by dense vegetation. Although further improvements are still needed to correct vegetation effects, nevertheless, the DENSE method improved the spatial resolution of SMAP soil moisture estimates, from 36 km to 1 km over the Tibetan plateau where high-resolution soil moisture products are necessary to support local and global hydrological applications.

Remote Sensing of Environment