InSAR-Derived Surface-Deformation-Constrained Machine Learning for Permafrost Occurrence Probability Mapping in the Eastern Kunlun Mountains

Accurate characterization of permafrost distribution on the Qinghai–Tibet Plateau is crucial for assessing permafrost degradation under climate warming and engineering disturbances, as well as its potential implications for cold-region infrastructure stability. However, statistical learning-based permafrost mapping is commonly constrained by the scarcity and uneven spatial distribution of reliable training samples. To mitigate this limitation, this study used Sentinel-1A SAR time series from 2017 to 2020 in the eastern Kunlun Mountains to derive two key deformation metrics: the mean annual surface deformation rate and seasonal deformation amplitude. Permafrost sample points were identified using thresholds on these deformation metrics under constraints related to glacier distribution, surface slope, and deformation time-series fitting quality. In parallel, non-permafrost sample points were obtained from areas consistently classified as non-permafrost by existing products under comparable spatial constraints. The selected sample points were used to train a gradient boosting machine (GBM) model, with mean annual air temperature (MAAT), scaled mean annual snow cover days (SMASCD), normalized difference vegetation index (NDVI), and slope as predictors, to generate permafrost occurrence probabilities across the study area. The results indicate that surface deformation in the study area is characterized by widespread subsidence and pronounced seasonal freeze–thaw deformation. The selected permafrost and non-permafrost samples showed clear separability in both deformation and environmental feature spaces, and the generated permafrost probability map indicated a permafrost extent of approximately 47,860 km2, accounting for 86.6% of the study area. Validation against 59 field-observed permafrost sites showed full consistency between the predicted permafrost occurrence and in situ observations. These findings indicate that the proposed method can support high-resolution permafrost mapping and degradation assessment in high-altitude regions.