The production of this dataset used a total of 1490 soil measurement data from sampling points. The main sources of the data are as follows: (1) The research team undertook the National Natural Science Foundation of China project "Study on the Gradient Variation Law and Mechanism of Soil Organic Carbon in Qilian Mountains" (41771252), "Investigation of Ecological Hydrological Sample Belt in Heihe River Basin" (91025002), "Simulation Study of Soil Carbon in Typical Ecosystems of Arid Inland River Basins" (31270482), the Gansu Provincial Major Project "Study on the Interaction between Qilian Mountain Water Conservation Ecosystem and Hydrological Processes and Their Adaptation to Climate Change" (18JR4RA002), and the Gansu Provincial Forestry and Grassland Bureau Science and Technology Innovation Project "Evaluation of Ecological Governance Effectiveness in Qilian Mountain National Park (Gansu Area)" (GYCX002). [2020] 01) Data from projects funded by the Gansu Provincial Science and Technology Plan, such as the "Gansu Qilian Mountain Ecological Environment Research Center" (18JR2RA026). (2) The original soil profile data and literature integration data (Xu et al., 2018 DOI: 10.1038/s41598-018-20764-9) publicly released by the National Glacial Permafrost Desert Science Data Center and the Qinghai Tibet Plateau Science Data Center, as well as the World Soil Reference and Information Center (ISRIC) WoSIS soil profile database（ https://www.isric.org/explore/wosis ）Wait. The raw data has undergone quality control to remove sampling points with latitude and longitude positioning accuracy less than 0.001, as well as outliers with significant deviation between soil organic carbon content and environmental background.

Based on the "Scorpion" (Soil, Climate, Organisms, Relief, Parent material, Age, Geographic position) framework, the spatial distribution of soil organic carbon density at a resolution of 30m in the 0-100 cm soil layer in the northwest region was simulated using the Digital Soil Mapping (DSM) method and operations based on tile structures. The prediction method is mainly based on the Extreme Gradient Boosting algorithm (XGBoost) in machine learning. Environmental covariate data includes Landsat 8 OLI multispectral images, Sentinel-1 radar images, temperature, precipitation, radiation, terrain, vegetation index, location and other raster data. Constructing a digital soil mapping framework in R language, in order to reduce memory usage and improve computing speed, the entire northwest region is divided into 21 sub regions of 600km × 500km. Within each sub region, a mapping process based on tiles (tile size: 45km × 45km) and parallel computing is constructed. The bootstrap method is used to repeat modeling, and spatial modeling is performed on each bootstrap sample to obtain the frequency distribution of the modeling results. The modeling uncertainty is represented by standard deviation (sd).

After 30 rounds of 10 fold cross validation, the RMSE and R2 of the model were 6.15 kg/m2 and 0.74, respectively. The final data product is divided into 21 sub regions, where mean and SD represent the mean and standard deviation of 30 repeated modeling, respectively, in kg/m2, representing the mass of soil organic carbon per unit area in the 0-100cm soil layer. After merging each sub region, the mean and standard deviation of soil organic carbon density in the entire northwest region are obtained, with a data volume of 63.70GB. The bilinear method is used to resample to 250m, and the total data volume of the final product is 435MB.