[Objective] A long-term and high-precision land cover dataset was constructed for the Loess Plateau. The spatiotemporal pattern of land cover in 2001 and 2020 was analyzed in order to provide a scientific underpinning for initiatives concerning ecological environmental preservation and sustainable development within the region. [Methods] Training samples were constructed using multiple sources of land cover products and ground feature data from various time periods. The Google Earth Engine (GEE) platform and a random forest classification model were used to generate the land cover of Loess Plateau (LCLP) dataset. Spatial analysis and a univariate linear regression model were then used to analyze the spatiotemporal pattern of land cover types on the Loess Plateau. [Results] According to the validation set built using random forest, LCLP exhibited an overall accuracy and kappa coefficient greater than 90%. Moreover, based on the independent verification set, LCLP demonstrated an overall accuracy ranging from 0.58% to 20.23% higher than existing products. Additionally, the accuracy of the classification of various land cover types, including cultivated land, forest land, grassland, impervious surface, and bare land, was increased. [Conclusion] Compared with other datasets, LCLP significantly improved classification accuracy and is suitable for accurately reflecting land cover changes for the Loess Plateau region. During 2001-2020, there has been a decreasing trend in cultivated land and shrubs in the Loess Plateau region, while forest land, water bodies, and impervious surfaces have shown a significant increasing trend. From the perspective of land cover changes, cultivated land and grassland were the primary sources of newly added land cover types.