Comparison of different downscaling schemes for obtaining regional high-resolution soil moisture data

• Two Soil Moisture (SM) downscaling schemes are compared over the Qinghai-Tibet Plateau (QTP). • Passive microwave brightness Temperature (TB) is downscaled to high resolution using machine learning methods. • High-resolution SM is retrieved from downscaled TB based on the radiative transfer model (RTM). • The “downscaling-then-retrieving” approach outperforms the “retrieving-then-downscaling” method over 70 % of the QTP. Obtaining regional fine-scale daily Soil Moisture (SM) data is crucial for better understanding carbon and water cycles. Currently, downscaling from passive microwave SM products has become the most commonly utilized approach for generating regional high-resolution SM data, while retrieving SM based on disaggregated brightness Temperature (TB) data gains litter attention. Besides, the relative potentials of these two downscaling approaches remains largely unknown. Therefore, this study comprehensively compared the relatively performances of the two downscaling schemes namely the “retrieving-then-downscaling” and “downscaling-then-retrieving” over the Qinghai-Tibet Plateau (QTP). Evaluation results showed that retrieving SM using disaggregated TB significantly outperformed the popular passive microwave SM downscaling method. The averaged R and ubRMSE metrics for downscale-first based SM were 0.62/0.74 and 0.051/0.038 m 3 /m 3 at station/network scales, and were 0.58/0.70 and 0.056/0.041 m 3 /m 3 for the retrieval-first based SM, respectively. Spatially, the downscale-first based SM had lower uncertainty than the retrieval-first based SM over nearly 70 % areas of the QTP. However, due to the strong vegetation scattering effect, it showed two times larger uncertainty than the retrieval-first based SM over densely vegetated regions in the east and southeast. In addition, satisfactory TB downscaling performance could be achieved by leveraging machine learning algorithms and multiple covariables, but need to further reduce additional errors. The superiority of “downscaling-then-retrieving” scheme was independent of models selected for downscaling. In conclusion, this study demonstrates the great potential of “downscaling-then-retrieving” method and provides a new insight for generating regional SM data at fine scale.