A multi-temporal sandy desert and sandy land classification dataset for the Mongolian Plateau from 1990 to 2020

Sandy desert (SD) and sandy land (SL) are distinct geomorphological types with different ecological and environmental significance. But they are often mixed in one land cover type (e.g., bare land) in the existing global and regional land cover datasets because of extraction method lacking. To address this challenge, we proposed a classification strategy that employs a Random Forest classifier, with integration of multi-source data combining spectral, topographic, and textural features. This approach separates SD and SL firstly, and applies the pixel dichotomy model to further subdivide SL into mobile, semi-mobile, semi-fixed, and fixed categories. Accuracy assessment demonstrates an overall accuracy (OA) of 92.39% and a kappa coefficient of 87.59%. The results show producer’s accuracy (PA) and user’s accuracy (UA) of 89.88% and 89.02% for SL, and 92.39% and 91.77% for SD, respectively. The dataset spans 1990–2020 at 5-year intervals, with a spatial resolution of 30 m. The resulting dataset provides big data support for Sustainable Development Goals 15.3—combating desertification and restoring degraded land in the fragile arid and semi-arid area.