Inter-comparison and integration of different soil moisture downscaling methods over the Qinghai-Tibet Plateau

•Six machine/deep learning methods (e.g. ANN, RF, GBDT, CNN, ResNet and LSTM) were used for downscaling and inter-compared.•A soil moisture downscaling method based on Machine learning and Three-Cornered Hat approach (MATCH) was proposed.•Distinct performance differences existed among diverse downscaling methods and the MATCH method showed great improvements. Soil moisture (SM) is a key state variable in the water, energy cycle between atmosphere and land surface but existing passive microwave soil moisture products typically have spatial resolutions of tens of kilometers, which fails to meet the requirements of regional applications. Even though numerous machine/deep learning methods have been applied to downscale SM, few studies have investigated the performance differences of diverse approaches and attempted to integrate various methods to improve downscaling accuracy. Therefore, this study firstly evaluated and inter-compared the downscaling performances of six machine/deep learning approaches and further proposed a hybrid downscaling method based on Bayesian three-cornered hat merging (MATCH). The daily 1 km seamless soil moisture product during 2015–2019 over the Qinghai-Tibet Plateau was then obtained. Evaluation and inter-comparison results revealed that there was obvious performance discrepancy among different downscaling approaches and the gradient boosting decision tree (GBDT) and random forest (RF) were the best two methods, which performed best in the southern and eastern of the plateau, respectively. While the artificial neural network (ANN) outperformed other approaches in the northwestern areas. Validation against in-situ measurements showed that compared with SMAP SM, the MATCH SM exhibited comparable accuracy and lower error with mean R and ubRMSE values of 0.55 and 0.047 m3/m3. The mean R and ubRMSE values for SMAP SM were 0.67 and 0.056 m3/m3, respectively. In addition, the MATCH SM presented great improvement compared with any single downscaled SM data, having the highest correlation and the lowest ubRMSE scores. While, among different methods, the highest R was 0.50 (GBDT) and the lowest ubRMSE was 0.052 m3/m3 (residual network (ResNet)). Besides, the downscaled SM could accurately reflect the temporal variations of soil moisture and precipitation, and effectively represent the spatial patterns of soil moisture. Satisfactory downscaling results were achieved in arid and semi-arid areas whereas a certain degree of overestimation still