As a large country with extensive grassland resources, China is facing severe grassland degradation. Studying trends in grassland vegetation cover change provides a basis for identifying the driving forces of grassland degradation and associated risk assessment. In previous studies, parametric regression models have typically been applied to estimate vegetation cover. However, the harsh assumptions of parametric regression have always been neglected. In the current study, vegetation cover monitoring data and vegetation indices (NDVI, SAVI, MSAVI, EVI), extracted from Landsat remote sensing images, were used to build random forest regressions, which are non-parametric models. These models were subsequently compared with traditional linear regression models. To build and test these models, 205 samples were collected from 2010 to 2015 (data from 2012 were not included) in alpine meadow, mountain meadow, lower-flat meadow, temperate meadow steppe, desert steppe, steppe desert, and desert in Burqin County, Xinjiang Uygur Autonomous Region. Among these samples, 150 samples were used to build models, and the remainder was used as testing data. Two sets of Landsat remote sensing images, Level 1 Standard Product and Surface Reflectance Product, were applied separately, and both included TM data for 2011-2012 and OLI data for 2013-2015. In total, two random forest models and 23 linear models were built. The results indicated that the predictive ability of the random forest models was clearly stronger than that of most of the linear models. Moreover, it was not necessary to test the assumptions for the random forest models, whereas none of the linear models' assumptions in this study could be satisfied perfectly. In the case study, random forest regression was applied to estimate the trend in grassland vegetation cover change in the last decade in the study area based on 663 sampling points. Among these, data for 2005-2009 were based on Landsat ETM+, data for 2010-2011 were based on Landsat TM, and data for 2013-2015 were based on Landsat OLI. It was clear that sensor differences would have a certain influence on the inversion results. Therefore, we also simultaneously built a random forest model for MODIS-EVI data, as this would not be affected by sensor differences, and the results calculated using MODIS data were considered to be a standard. For Level 1 standard data, the results based on Landsat ETM+ were significantly smaller than the results based on MODIS data. For surface reflectance data, the influence of different sensors on the results could be markedly reduced. Finally, to quantify the uncertainty of vegetation cover change trend based on surface reflectance data, we used a random forest model to verify vegetation cover extracted from different sensors during the same period, and found that the uncertainty was between -10% and 11%. In conclusion, random forest regression is assumed to be a better model to inverse vegetation cover than linear models. For its application in time series studies, Landsat surface reflectance production could significantly reduce the influence of sensor difference, although the uncertainty was still approximately ±10%. In the current study, we assessed many grassland types, and to improve the accuracy of prediction, vegetation indexes should be calculated separately in further studies. In addition, efforts should be made to reduce the uncertainty associated with the data from different types of sensor.