Integrating optical and InSAR data for machine learning-based rock glacier activity classification: Model evaluation across mountain regions

Rock glaciers (RGs) are crucial geomorphic indicators of mountain permafrost. While extensive RG inventories exist, determining RG activity status (i.e. intact and relict) has traditionally relied on field surveys and visual interpretation. This study utilizes high-resolution Planet satellite imagery to automatically classify the activity status (intact vs. relict) of pre-mapped rock glacier polygons. To achieve this, we implemented a hierarchical pipeline: (1) extraction of image patches based on inventory polygons; (2) patch-level inference using machine learning models; and (3) aggregation of predictions via majority voting to determine the final status of each RG. Six machine learning algorithms – including random forest (RF), residual neural network (ResNet), k-nearest neighbor (KNN), decision tree (DT), logistic regression (LR), and support vector machine (SVM) – were evaluated based on RG samples mapped using geomorphic interpretation and interferometric synthetic aperture radar (InSAR) surface displacements in Daxue Shan, Southeastern Tibetan Plateau. The training results show that the RF classifier exhibited superior performance among the classifiers, achieving an accuracy of 86.1%. Validations in the Nepalese Himalaya and Gangdise Mountains confirmed the transferability and reliability of the RF model, with accuracies of 85.48% and 90.91%, respectively. Analysis of the relationship between the classified intact and relict RGs and the global permafrost zonation index (PZI) map revealed a strong correlation between the prevalence of intact RGs and permafrost-prone areas. However, the statistical distribution of relict RGs in both test regions highlights a scale-dependent mismatch between coarse-resolution permafrost models and local landform conditions. This underscores the value of RG activity status as complementary local evidence to refine and validate future permafrost distribution maps.