Long-term Land Cover Dataset of the Mongolian Plateau Based on Multi-source Data and Rich Sample Annotations

The Mongolian Plateau (MP), with its unique geographical landscape and nomadic cultural features, plays an important role in the safety of the regional ecological environment and sustainable development in North Asia. Dramatic land cover changes in the MP cannot be tracked by existing global-scale land cover datasets because existing global land cover products often lack the classification specificity and temporal continuity required for MP-specific studies, particularly for grassland and bare area subtypes. To address this gap, a new land cover classification was designed for MP, which includes forests, shrubs, meadows, real steppes, dry steppes, desert steppes, wetlands, water, croplands, built-up land, barren land, desert, sand, and ice. By leveraging machine learning and cloud computing, we developed a novel dataset spanning the period of 1990–2020. This involved manually annotating 64,345 sample points within a 0.8°×0.5° gridded framework across the MP and training a random forest algorithm with multi-source feature bands (e.g., spectral bands of Landsat, NDVI, NDWI, elevation, and building volume) to produce annual land-cover maps at five-year intervals. The resulting dataset was validated with an overall accuracy of 83.6% (kappa = 0.807). The detailed subclasses of grassland, specific desert, and sand types are highlighted based on maintaining maximal consistency with other global-scale land cover datasets. The key findings include the dominance of real steppe, barren land, sand, and forest across the MP, with significant trends over the 30-year period; real steppe and cropland areas increased (4,317 km²/year and 1,516 km²/year, respectively), barren land and desert areas decreased (-6,665 km²/year and -496 km²/year, respectively), and sand areas expanded dramatically (1,372 km²/year). Over the past 30 years, the MP has experienced degradation followed by recovery, with a turning point in 2005. The recovery was mainly characterized by the transition from dry steppe to real steppe and from barren land to desert steppe. Thus, this product provides a comprehensive record of land cover dynamics, supported by seven periods of land cover data and rich labelled samples. It has high potential for application in rangeland resource management and regional sustainable development in vast arid and semi-arid areas.