3 Methodology
The method we adopted consists of two parts and is detailed below. First, we mapped active rock glaciers manually from interferograms and Google Earth images. Second, we used the manually labelled images to train a deep learning network, i.e., DeepLabv3+, for mapping rock glaciers automatically from Sentinel-2 optical images.
3.1 Mapping active rock glaciers from interferograms and Google Earth images
In this subsection, we first describe the strategy of delineating rock glaciers. Then we present the method for quantifying rock glacier kinematics by InSAR. Finally, we introduce how to determine the geomorphic attributes of the mapped landforms.
3.1.1 Manual identification and delineation of rock glaciers
We mapped active rock glaciers by combining two imagery sources: interferograms and Google Earth images. The displacement maps generated by InSAR allow us to easily recognize moving parts of the ground surface, meanwhile the high-resolution and multi-temporal Google Earth images provide geomorphic information to distinguish rock glaciers from the other active surface units, such as debris-covered glaciers, solifluction lobes, and slow-moving landslides. Visual identification was conducted based on the geomorphological criteria proposed by RGIK (2021) including the frontal and lateral margin morphology, and the surface ridge-and-furrow topography as an optional indicator. We then outlined the recognized landforms along their extended geomorphological footprints, i.e., the frontal and lateral margins are included within the boundaries. We followed the IPA guidelines because it provides practical and standardized baseline concepts for identifying and outlining rock glaciers from remote sensing images and readily applicable to producing consistent inventories over wide-extent regions.
3.1.2 Kinematic quantification by InSAR
In total, twenty-two interferograms generated from ALOS-1 PALSAR images covering the West Kunlun Mountains were used for ground movement detection between 2007–2008 (Table 1). To maintain high interferometric coherence and reduce topographic error, we selected image pairs with temporal spans of 46 days and perpendicular baselines smaller than 1,000 m. The topographic phase were estimated and removed by using a digital elevation model (DEM) produced by the Shuttle Radar Topography Mission (SRTM) with a spatial resolution of ~30 m over most of the study region. A tile of TanDEM-X DEM (spatial resolution ~12 m) was adopted for correcting topographic phases for one interferogram overlapping with the permafrost survey region. Multi-looking operation and adaptive Goldstein filter (8\(\times\)8 pixels) were applied in the interferometric processing, which was implemented by the open-source software InSAR Scientific Computing Environment (ISCE) version 2.2.0 (available athttps://github.com/isce-framework/isce2 ). We then unwrapped the interferograms with the SNAPHU (Chen and Zebker 2002) and selected one point located at the flat and stable ground close to each rock glacier to re-reference the unwrapped phases measured within the boundary of each landform. By doing so, we managed to remove the long-wavelength orbital errors and the atmospheric artefacts including the water vapor delay and ionospheric effects, all of which can be assumed identical within the extent of a rock glacier (Hanssen 2001).
We determined the surface downslope velocities of rock glaciers as their kinematic attributes. The surface velocities along the SAR satellite line-of-sight (LOS) direction were derived from the unwrapped interferograms and then projected to the downslope direction of each landform (Hu et al. 2021). Associated uncertainties including the InSAR measurements and geometric parameters were quantified through error propagation (Hu et al. 2021). We used the spatial mean velocity within a rock glacier to represent its overall kinematic status. Then we refined the results by selecting data that fulfilled the following criteria: (1) after masking out the pixels with low coherence (< 0.3) (Wang et al. 2017), the remaining pixels account for more than 40% of the entire landform extent; (2) the relative errors of the spatial mean velocities are lower than 20%.

3.1.3 Determination of geomorphic attributes

Essential geomorphic attributes such as the elevation range, mean slope angle, and landform aspect were quantified using the SRTM DEM. Qualitative attributes including the spatial connection of the rock glacier to the upslope unit and the activity category were described and assigned to the dataset following the IPA guideline (RGIK, 2021). We primarily classified the mapped rock glaciers according to their spatial connection to the upslope unit because it could provide implications regarding the landform genesis (Sect. 5.2). We used the Global Land Ice Measurements from Space (GLIMS) dataset to help recognize the surrounding glacier units (GLIMS and NSIDC, 2005). Figure 2 presents examples of rock glaciers that were classified by their upslope units into four categories. For instance, Figure 2b shows a glacier-connected rock glacier, the frontal and lateral margins of which are discernible from the Google Earth image, though the rooting zone is ambiguous. We separated the rock glacier from the upslope unit from surface structure in this case. Finally, we created the InSAR-based sub-dataset. The entire workflow is illustrated in Figure 3 with one example shown in Figure 4.