Plastic film mulching has greatly increased crop yields in arid and semi-arid regions of China, but also caused a lot of environmental problems. Thus, timely and accurate mapping of plastic-mulched farmlands through remote sensing technology is helpful for governments to plan agricultural production and deal with micro-plastic pollutions. However, the existing recognition methods based on single-temporal remote-sensing images with low and medium resolutions are unable to accurately recognize the plastic-mulched farmlands in the Loess Plateau due to its complex terrain and fragmented agricultural landscapes. In order to accurately recognize plastic-mulched farmlands in the Loess Plateau, different feature set combination schemes and multi-temporal image combination schemes applicable to recognize plastic-mulched farmlands in the Loess Plateau were constructed based on Sentinel-2 remote-sensing images and random forest algorithm. Three testing areas were selected for constructing recognition schemes mentioned above, including Linxia County in Gansu Province, Pengyang County in Ningxia Hui Autonomous Region, and Shanyin County in Shanxi Province, and one validation area of Xunyi County in Shaanxi Province were chosen as the scheme validation area. Firstly, based on the random forest algorithm, the optimal feature set combination scheme with the highest recognition accuracy was selected from seven different feature set combination schemes for each growth stage (sowing stage, flourishing stage, and harvesting stage). Then, based on the remotesensing images of the three different crop growth stages and their corresponding optimal feature set combination schemes, different multi-temporal image combination schemes were constructed to recognize the plastic-mulched farmlands, and then the optimal multi-temporal image combination scheme was selected. Finally, the effectiveness of the optimal feature set combination scheme and multi-temporal image combination sheme for recognizing plastic-mulched farmlands was verified in Xunyi County, and the spatial distribution maps of plastic-mulched farmland in each research area were drawn. The results showed that the visible bands (B2, B3, and B4) and the short-wave infrared bands (B11 and B12) in the spectral feature set of Sentinel-2 remote-sensing images, the normalized difference bareness and built-up index (NDBBI), normalized difference water index (NDWI), bare soil index (BSI), normalized difference built-up index (NDBI), and modified normalized difference water index (MNDWI) in the index feature set, and the sum average (savg) and correlation (corr) in the textural feature set can be used as optimal input feature variables for recognizing plastic-mulched farmlands. Among the seven different feature set combination schemes, the “spectum + index” scheme was the optimal scheme for recognizing plastic-mulched farmlands during the sowing and harvesting stages. The F1-score for plastic-mulched farmland recognition in these two stages in the four study areas was greater than 87% and 57%, respectively. The “spctrum + index + texture” scheme was the optimal scheme for recognizing plastic-mulched farmlands during the flourishing stage with F1-score greater than 71% in the four study areas. Generally, the plastic-mulched farmland recognition accuracy based on multi-temporal remote-sensing images was higher than that based on single-temporal remote-sensing images. Among different multi-temporal image combination schemes, “sowing stage + flourishing stage + harvesting stage” can be used as the optimal scheme for plasticmulched farmland recognition, and the F1-score for recognizing plastic-mulched farmlands in the four study areas was greater than 92%. In general, plastic-mulched farmlands in the Loess Plateau can be accurately recognized based on random forest algorithm and the optimal feature set combination schemes and multi-temporal image combination scheme.