AbstractMachine learning combined with multisource remote sensing data to assess soil moisture content (SMC) has attracted considerable attention in SMC studies, but the retrieval results still remain uncertain. The purpose of this study is to combine multiple single machine learning models with integrated learning algorithms and propose an SMC retrieval method based on multiple differentiated models under a stacking integrated learning architecture. First, 19 factors, including: radar backscattering coefficient, vegetation index, and drought index, that affect SMC were extracted from SENTINEL‐1, LANDSAT, and terrain factors. Those with the highest importance scores were selected as retrieval factors using the Boruta algorithm combined with four single machine learning methods—classified regression tree, random forest, gradient boosting decision tree (GBDT), and extreme random tree. In addition, the two stacking ensemble models using least absolute shrinkage and selection operator (LASSO) and the generalized boosted regression model (GBM) were tested and applied to build the most reliable and accurate estimation model. The results showed that radar backscattering coefficient, temperature, vegetation drought index, land surface temperature, enhanced vegetation index, and solar local incident angle were the most important environmental variables for soil moisture retrieval. A comparison of the four machine learning methods in April and August showed that the GBDT model revealed the highest SMC retrieval accuracy, with root mean square error values of 1.87% and 1.64%, respectively. The stacking models were more accurate than the optimal single machine learning model, especially when using GBM. The multifactor integrated model constructed using spectral indices, radar backscatter coefficients, and topographic data exhibited high accuracy in soil surface moisture retrieval in an arid zone, providing a reference for land desertification studies and ecological environment management in the study region.