Downscaling SMAP Soil Moisture Using a Combination of XGBoost and DSCGAN Algorithms
Due to the coupling effects of climatic conditions,surface and subsurface conditions,and human activities,soil moisture is highly heterogeneous on spatial and temporal scales.The SMAP soil moisture products from satellite microwave remote sensing can be used from continental to global scales,but they are not suitable for small-and medium-scale applications due to low spatial resolution.To improve the spatial resolution of soil moisture products,various downscaling methods have been developed,with the empirical downscaling method being widely used due to its relatively simple calculation.These models require downscaling factors,which are mostly obtained based on optical remote sensing and are susceptible to cloud influence.Therefore,it is impossible to obtain high spatial resolution soil moisture continuously over time using this model for downscaling.To solve this problem,we proposed a downscaling framework based on multiple data sources using machine learning and deep learning methods.The main idea is to use traditional machine learning methods in the absence of clouds and super-resolution methods to downscale soil moisture in the presence of clouds.The combination of these two methods yields time-continuous,high-resolution soil moisture.First,multi-source data were used to obtain fifteen downscaling factors,including surface temperature,normalized vegetation index,albedo,elevation,slope,slope direction,soil cover type,soil texture,etc.Then,three machine learning models(Random Forest,LightGBM,and XGBoost)were used to establish empirical downscaling models of SMAP soil moisture product data with downscaling factors.The best performing XGBoost model was chosen to downscale the spatial resolution of SMAP soil moisture products from 9 km to 1 km.Finally,the DSCGAN super-resolution model was trained based on 9 km and 1 km soil moisture data pairs.The trained models were used to obtain spatio-temporally continuous soil moisture data for the study area.The results show that,by comparing the downscaled soil moisture and original SMAP data,the R is 0.96,the RMSE is 0.034 m3/m3,the bias is 0.011 m3/m3,and the ubRMSE is 0.034 m3/m3.The comparison between the downscaled soil moisture and the measured data shows that the R is 0.696,the RMSE is 0.192 m3/m,the bias is-0.171 m3/m3,and the ubRMSE is 0.089 m3/m3.The downscaling method proposed in this study provides a framework for generating higher resolution spatio-temporally continuous surface soil moisture that can meet the needs of small-scale applications such as regional moisture surveys and agricultural drought monitoring.
soil moisture downscalingSMAPmachine learningdeep learningShandian river basin
万方数据
维普
