[A-CC32] Glaciology

[ACC32-11] Machine learning-based spatial analysis of the spring states in the southernmost Eurasian permafrost, Hangai Mountains, central Mongolia

*Ishikawa Mamoru¹, Tetsuya Hiyama², Khurelbaatar Temuujin³, Dashsteren Avirmed ³ (1.Hokkaido University, 2.Nagoya University, 3.Institute of Geography and Geoecology, Mongolian Academy of Science)

Keywords:Machine learning, Mongolia, Permafrost, Spring discharge

The springs are the crucial ecosystem resource for pastoralism livelihood and are recently under depleting in the southernmost regions of Eurasian permafrost. This study aims to evaluate recent spring states (either still discharging or already depleted) that were discharging at the time of several decades ago in the Hangai Mountains, Mongolia. There are a total of 1620 spring sites, among which 228 sites were surveyed by field visits in July and August 2019, and analyzing the visible satellite images taken between 2008 and 2020. We predicted the other spring states, using machine learning approaches as logistic regression (LR), random forest (RF) and support vector machine (SVM), and explanatory variables of vegetation and topography-derived hydrological parameters. Owing to imbalanced (depleted: 44, discharging: 192) and small number of training datasets, we applied cross-validation and min-max normalization to the modelling to minimize overfitting issues. The resulted models showed high prediction performance as in area under curves of receiver operating characteristics (AUC-ROC) are 0.84 (LR), 0.86 (RF), and 0.84 (SVM). Also, AUC-PR (AUC of Precision and Recall), criteria indicating model performance for imbalanced datasets, increased by 0.32 to 0.52 from the case in random classification. The most important explanatory variables to determine spring states were modified normalized difference water index (MNWDI) and normalized different vegetation index (NDVI); the springs in the relatively wet and green environments tend to remain discharging, while those in arid are depleted. The ensemble of three models predicted that 22.5 % of springs have already depleted in the Hangai Mountains. In widespread permafrost areas, the springs are depleted due to decreases in modern precipitation. Where permafrost and non-permafrost areas coexist, thickening of active layer and increases of unfrozen water contents contribute to remaining springs discharging despite ongoing aridification.