Mapping of peatland in Mongolia using machine learning

Mongolia ranks among the top 10 countries by peatland area and represents an important global carbon sink. However, the distribution of its peatlands remains largely unconfirmed. Here, we combine field survey and remote sensing data to improve estimates of peatland distribution by modeling and mapping soil organic carbon content (SOCC). The case study dataset comprises 5400 soil samples from 1242 sites collected between 2000 and 2023, spanning Mongolia's major natural zones. Peat soils were defined using a threshold of >120 g kg −1 SOCC for the upper 0–30 cm of the soil profile. To predict SOCC, we evaluated thirty-five environmental covariates, including climate data, optical and thermal Landsat-8 imagery, vegetation indices, terrain factors, and soil properties. We implemented a random forest model within a spatial prediction framework, training it on the field soil observations. Cross-validation showed that the SOCC model achieved high accuracy ( r = 0.89 ), with the most important predictors being temperature, silt content, vegetation productivity, and moisture availability. We map a peatland area of 47,050 km 2 , equivalent to 2.9% of Mongolia's territory, concentrated in permafrost high-mountain, taiga, and forest-steppe zones of the major northern and western mountain ranges. Unlike previous estimates, which relied on coarse extrapolation or unvalidated inventories, our approach integrates an extensive field dataset with a machine learning method to produce the first detailed, field-validated national peatland map for Mongolia at 250 m resolution. Our work resolves fine-scale spatial patterns of carbon storage while quantifying predictor importance, providing a critical dataset and ecological indicator for evaluating carbon sinks and sources and for conservation planning in this climate-sensitive region.