A total of 973 soil sampling points in the Qilian Mountains and surrounding areas were used to produce this dataset, with the following main sources of data

Projects undertaken by this research team：

The National Natural Science Foundation of China (NSFC) project "Research on the variation pattern and mechanism of soil organic carbon along the gradient in Qilian Mountain" (41771252), "Investigation of Ecohydrological Sample Strips in the Heihe River Basin" (91025002)、"Soil Carbon Simulation in Typical Ecosystems of Arid Inland River Basins" (31270482); Major Project in Gansu Province: "Research on the Interaction between Ecosystems and Hydrological Processes in Qilian Mountains and Their Adaptation to Climate Change" (18JR4RA002), The Science and Technology Innovation Project of Gansu Provincial Forestry and Grassland Bureau, "Research on Ecological Management Effectiveness Assessment of Qilian Mountain National Park (Gansu Area)" (GYCX[2020]01), Gansu Provincial Science and Technology Plan Funding Project "Gansu Qilian Mountain Ecological Environment Research Centre" ( 18JR2RA026) and other project data.

Raw Soil Profile Data, Literature Integration Data published by National Cryosphere Desert Data Centre and National Tibetan Plateau Data Center (Xu et al., 2018 DOI:10.1038/s41598-018-20764-9), WoSIS Soil Profile Database published by International Soil Reference and Information (ISRIC, https:// www.isric.org/explore/wosis), among others.

The original data were quality controlled by removing sampling points with less than 0.001 latitude and longitude positioning accuracy, and abnormal values where the soil organic carbon content deviated significantly from the environmental background.

Based on the 'Scorpan' (Soils, Climate, Organisms, Relief, Parent material, Age, Geographic position) framework, the spatial distribution of soil organic carbon density at 30m resolution in the 0-100 cm soil layer was simulated using the Digital Soil Mapping (DSM) method and a tile structure-based algorithm.

The forecasting method is mainly based on the Extreme Gradient Boosting algorithm in machine learning, with environmental covariate data including: Landsat8 OLI multispectral imagery, Sentinel-1 radar imagery, raster data such as temperature, precipitation, radiation, topography, vegetation indices, and location.

A framework of tile-based and parallel computing mapping is constructed in R. The modelling is repeated by the bootstrap method, and the spatial modelling is performed for each bootstrap sample to obtain the frequency distribution of the modelling results, with the modelling uncertainty expressed as standard deviation (sd).