Machine-learning-based spatial analysis of the spring states in the southernmost Eurasian permafrost, Hangai Mountains, central Mongolia

Abstract Springs are the vital ecosystem resource for pastoral livelihood in the southernmost regions of Eurasian permafrost but are currently experiencing depletion. However, their spatial distribution and factors influencing this depletion remain unassessed. This study evaluates the recent status of springs that were discharging several decades ago in the Hangai Mountains, Mongolia. A total of 1,620 spring sites were identified, and states (either still discharging or already depleted) of 228 springs were determined through field observations conducted in July and August 2019, along with visible satellite imagery taken between 2008 and 2020. To predict the states of remaining 1392 springs, machine learning approaches, including logistic regression (LR), random forest (RF), and support vector machine (SVM), were applied, incorporating vegetation indices and topographically derived hydrological parameters as explanatory variables. Given the imbalanced datasets (depleted: 44, discharging: 192) and the small sample size of training and test datasets, cross-validation and min–max normalization were employed to minimize overfitting problems. The models demonstrated high predictive performance, with area under curves of receiver operating characteristics (AUC-ROC) are 0.84 (LR), 0.86 (RF), and 0.84 (SVM). Additionally, the AUC of precision and recall (AUC-PR), criteria indicating the model performance for imbalanced datasets, increased by 0.32–0.52 compared to the case in random classification. The most important explanatory variables for determining spring states were the Modified Normalized Difference Water Index (MNDWI) and the Normalized Different Vegetation Index (NDVI); springs in the relatively wet and vegetated environments tended to remain discharging, whereas those in arid environments were more likely to be depleted. An ensemble of the three models predicted that 22.5% of springs in the Hangai Mountains have already been depleted. In extensive permafrost regions, spring depletion is largely driven by decline in modern precipitation. In transitional zones where permafrost and non-permafrost areas coexist, the thickening of active layer and increases in unfrozen water contents help to sustain spring discharge despite ongoing aridification.