Shrubs are important ecological barriers in desert regions and an important component of global carbon estimation. However, the shrubland in deserts has been hardly presented, although many high-quality land cover datasets with a 10 m scale based on remote-sensing data have been publicly released products. Therefore, the underestimation of carbon storage is inevitable with the absence of desert shrublands. The existing land-cover datasets have been analyzed and compared, and it has been found that the reason for missing the shrubland in deserts is mainly indued by the absence of shrubland samples, which are easy to neglect and difficult to retrieve. In this study, we developed a semi-automatic method to extract shrubland samples in deserts as the updated input for the machine-learning method. Firstly, the initial samples of desert shrublands were identified from the very high spatial-resolution (0.3~0.5 m) imagery on GEE, and the maximum NDVI from Sentinel-2 was used for double-checking. Secondly, a feature-based method was used to learn the feature from the initial samples and a similarity-based searching method was employed to automatically expand the samples. Finally, the expanded samples and their corresponding time-series satellite images were inputted into different machine-learning methods at a large region (1.63 × 10⁶ km²) for extracting the shrubland in the desert. It was found that different combinations of feature variables and time-series combinations have different impacts on the overall accuracy (OA) of the classification results, as well as the performance of identifying and classifying the different land-cover types. Compared to the existing global-scale land-cover products, the proposed method can better identify the shrubland in deserts and show better overall accuracy.