Figure 12. Histograms of the downslope velocities for (a) all RGs, (b) DMS-RGs, (c) G-RGs, (d) GF-RGs, and (e) T-RGs, respectively.
5 Discussion
In this section, we firstly summarize the potential and limitations of using the combined methodology for mapping rock glaciers (Sect. 5.1). Then we discuss the genetic and evolutional implications carried by the geomorphic characteristics of the mapped rock glaciers (Sect. 5.2).
5.1 Potential and limitations of the InSAR-Deep learning combined method for mapping rock glaciers
We used an InSAR-Deep learning combined approach to map rock glaciers across the West Kunlun Mountains. The advantage of the combined methodology is twofold: the InSAR-based mapping approach provides essential information on surface kinematics and accurate manual delineation for training the deep learning model; whereas the automated method improves mapping efficiency and more importantly, overcomes the conservativeness of the former approach and expands the InSAR-based sub-dataset. More specifically, some rock glaciers cannot be detected by InSAR due to coherence loss in interferogram, geometric distortions, their topographic orientations insensitive to InSAR line-of-sight measurements, or simply their inactive kinematic status (Wang et al. 2017; Robson et al. 2020). As we used the conventional Differential InSAR method, the smaller amount of interferograms adopted for identifying rock glaciers could lead to more serious omission in the dataset compared with using multi-temporal data (Cai et al., 2021; Zhang et al., 2021; Bertone et al., 2022). By combining the deep learning method, we can map the landforms that had been omitted due to coherence loss in the limited number of interferograms. In addition, rock glaciers moving parallel to the satellite direction, or along a steep slope, or at a very fast or slow pace, can be mapped as well.
However, our deep learning approach has a limited level of automation: the results produced by this methodology still requires manual inspections and modifications to increase the accuracy. Among the factors controlling the deep learning performance, the amount and quality of training and validation samples is one primary factor that affects the mapping accuracy. In this study, the training and validation datasets consist of the boundaries of active rock glaciers in the InSAR-based sub-dataset overlying the Sentinel-2 optical images (examples as shown in Figure 5). The amount of rock glaciers (172) as training and validation samples is in the same order of magnitude as the landform amount (338) used by Robson et al. (2020) for training their deep learning network; yet the training data size can be improved to fully achieve the potential of the state-of-the-art network (DeepLabv3+Xception71) we adopted. Quality of the input images is also moderate, as the Sentinel-2 images have a medium spatial resolution of ~10 m, making it challenging to characterize some rock glaciers, especially small ones with areas smaller than 30,000 m2, from these optical images and possibly leading to inaccuracy in the output. Therefore, manual inspection is required in the post-processing to improve the accuracy of the automatically delineated boundaries. Additionally, the cloud cover of the images hinders the compilation of a complete inventory across the large area. Finally, the Google Earth images (2009–2020) we referred to while creating the InSAR-based sub-dataset are unsynchronized with the Sentinel-2 images (Jul–Aug of 2018) used for producing the training data and for predicting rock glaciers by the trained model. Accordingly, we conducted additional manual inspections while preparing the input data and recognized few differences requiring corrections to the training data because the rock glacier activity is relatively low in the study area (Sect. 4.3), yet this asynchronization may lead to errors in areas where rock glaciers have been moving fast in recent decades.
Furthermore, as we evaluated the effectiveness of the deep learning-based method by applying the trained model to a test area outside the original study area and the validation IoU, which reached a value of ~0.8 comparable with the previous milestone research (Chen et al., 2018), the imperfect metric we achieved (i.e., validation IoU < 1) reveals the possibility that some rock glaciers may still be missed in our inventory. We estimated the magnitude of landform underestimation by calculating an index from the validation IoU and a test experiment in a new region (methodology detailed in Text S1); yet it is challenging to provide a precise estimate given that no ground truth data is available over the study region.
In addition, our combined approach is limited to mapping intact landforms, i.e., active and transitional rock glaciers according to the updated categorization scheme of rock glacier activity proposed by RGIK (2021). The InSAR-based sub-inventory contains active rock glaciers, the surface of which display coherent downslope motion as revealed by the interferograms. The transitional rock glaciers, on the other hand, show little movement over the surface, yet their geomorphologic characteristics are less distinguishable from the active landforms. Our deep learning model essentially learned the visual features of active rock glaciers through the optical images in the training dataset, and thus the model is likely to identify and delineate transitional rock glaciers as well. In contrast, relict rock glaciers usually develop distinct geomorphologic features such as subdued topography and vegetation cover, which cannot be mapped by the deep-learning model.
Considering the above limitations, several improvements can be implemented in our future research: (1) to increase the amount and diversity of training samples by including rock glacier boundaries from other regions; (2) to adopt higher-resolution and more cloud-free optical images for producing input dataset; and (3) to use generative adversarial network for translating optical images (for landform inference) to the domain of training data and include them during training. Nevertheless, the developed model will be useful for regions where data gap exists, such as many mountain ranges on the Tibetan Plateau. The inventory produced by this work will serve as an important database for scientific investigations such as managing geohazards (e.g., Kummert and Delaloye, 2018), assessing sediment budget (e.g., Kofler et al., 2022), and monitoring permafrost changes (e.g., Thibert and Bodin, 2022).
5.2 Genetic and evolutional implications from the geomorphic characteristics of rock glaciers
We classified the mapped rock glaciers into glacier-connected (G-RGs), glacier-forefield-connected (GF-RGs), debris-mantled slope-connected (DMS-RGs), and talus-connected rock glaciers (T-RGs). This classification scheme was adopted firstly for a practical reason: spatial connection of the rock glacier to its upslope unit is mostly well discernible from the optical images (as illustrated in Figure 2). Moreover, we take the distinction as an indication of the evolution of rock glaciers in terms of their ice origin, sediment source, and debris transfer process. In this subsection, we interpret the genetic and evolutional implications held by the characteristics of rock glaciers in the regional geomorphologic context.
Nearly half (~49%) of the mapped rock glaciers are spatially connected to glaciers. The amount appears to be reasonable because much of the West Kunlun Mountains (~12,500 km2) is occupied by modern glaciers (Kääb et al. 2015), constituting one of the most prominent glacierization centers on the Tibetan Plateau (Shi 2006). G-RGs occurring at the immediate downslope of the modern glaciers are likely to have the ice core embedded within the landforms, representing the transitional process from glacier (or debris-covered glacier) to rock glacier (Potter 1972; Whalley and Azizi 1994). However, we postulate that such transition is not actively ongoing given that glaciers in the West Kunlun are in mass balance or even slightly gaining mass in recent decades (Bao et al. 2015; Kääb et al. 2015; Wang et al. 2018; Zhou et al. 2018). The G-RGs are likely to gradually evolve from glaciers since the last cold period, i.e., the Little Ice Age (LIA, 200–600 aBP), and this transitional process tends to slow down in the past several decades (Shi 2006).
Although the landform transition is currently not active in our study area, we propose that the glacier-to-rock glacier continuum, as one classical theory about rock glacier genesis (Berthling 2011), can be adopted to interpret the evolution of the GF-RGs in our inventory. The GF-RGs are spatially disconnected from the upslope modern glaciers (Figure 2c), occurring at the lowest altitudes among all the categories in the study area (Sect. 4.2). Interactions between the GF-RGs and the glacier units are likely to take place during the glacier advance phases in geologic history. Anderson et al. (2018) modelled the glacier – debris-covered glacier – rock glacier evolutional process by simulating the rise of environmental equilibrium line altitude in response to climate warming: a pure glacier melts and separates from the emerging debris-covered terminus, which preserves its ice core due to insulation effect produced by the surface sediment and finally transforms into a rock glacier. Accordingly, we postulate that the GF-RGs in our study area once were part of the upslope glaciers during the Neoglaciation (3000–4000 aBP), when glaciers extended to altitudes hundreds of meters lower than the present glacier termini in the West Kunlun (Li and Shi 1992; Shi 2006).
Interactions between glaciers and rock glaciers are highlighted in the West Kunlun Mountains by the occurrence of abundant surge-type glaciers, whose flow velocities peaking at 0.2–1 km yr-1 during their active phases (Quincey et al. 2015; Yasuda and Furuya 2015). Excess materials consisting of ice and debris are carried downslope to areas far beyond the normal termini of the surge-type glaciers and may deliver sediments to the nearby glaciers (or debris-covered glaciers), whereby the surge events tend to contribute to the glacier-to-rock glacier transition provided that glaciers in the West Kunlun will retreat in the future as the glaciers in other alpine regions worldwide nowadays. A comparable case is the ongoing glacier-to-rock glacier transition in the Himalayas: based on field observation and sedimentologic analysis, Jones et al. (2019) elaborated that debris supply from the environmental sediment sources (in addition to the sediment derived from glaciation of the transitional landform per se) drives the evolution as an important factor. Moreover, the ice-debris body transferred and deposited at the far end of a surge-type glacier may gradually evolve into a rock glacier under favorable climatic and topographic conditions. Figure 13 presents an example of the potential evolutional process: the terminus of a surge-type glacier is covered by debris (Chudley and Willis, 2019) and many thermokarst lakes develop on the surface of the ice-debris mixture, which is likely to transform into a rock glacier in a warming climate. Two rock glaciers (wkl019 and wkl020) are situated in the surroundings and may receive debris and ice input during the surge events.
The genesis of two categories of rock glaciers, namely the T-RGs and the DMS-RGs, are related to periglacial processes. The T-RGs are conventionally considered as features originated in the periglacial domain: the rock glaciers contain interstitial ice developed by various processes such as burial of surface snow that typically occur in the formation of frozen ground (Humlum 1988; Haeberli 2000; Berthling 2011). The DMS-RGs are seldomly reported in the literature (Hu et al. 2021), yet display unique geomorphologic characteristics and constitute the second largest category (~35%) in the study area. In the absence of an upslope glacial system, we suggest that the DMS-RGs also represent the periglacial processes controlling the landform genesis. In comparison with the other three categories, the DMS-RGs occupy the highest and steepest slopes, where mechanical weathering dominates and produces sufficient sediments transferred and accumulated to the base of the slopes. During the glacial period, interstitial ice is formed within the deposits. The ice-debris mixture gradually develops and at one point overcomes the friction and starts to creep as an active rock glacier. Considering the lack of a headwall and the very small dimension (~one fifth of the average size of all mapped landforms, 0.05 km2 vs. 0.26 km2), it is likely that the DMS-RGs began to emerge during the Little Ice Age and are still at their embryonic stage.