Figure 5 . Diagram of the workflow to automatically map rock glaciers using DeepLabv3+ network. AI stands for artificial intelligence.
4 Results
We compiled an inventory consisting of 413 rock glaciers across the West Kunlun Mountains: 290 of them were mapped by the conventional method based on interferograms and Google Earth images, the other 123 landforms were identified by deep learning network with supplementary modifications to the automatically delineated boundaries (Figure 1).
In this section, we first present the accuracy of the automated mapping method. Then we analyze the features of all the mapped rock glaciers from the geomorphological perspective. Finally, we summarize the kinematic characteristics of the active rock glaciers measured by InSAR.
4.1 Performance of the automated mapping approach
After iteratively training and improving the model (Sect. 3.2), we trained a model attaining a performance of IoU = 0.801 on both the training and validation datasets (Figure 6).
Over the entire West Kunlun region, our trained model automatically identified and delineated 337 landforms as rock glaciers, among which 123 rock glaciers were newly discovered, 49 predicted polygons were false positives, the rest (165) were true positives but already present in the InSAR-based sub-dataset. Figure 7a and b present the satisfactory accuracy of automated delineation by comparing the deep learning mapped rock glaciers with the manually mapped boundaries in the training and validation datasets, respectively. And Figure 7b is an example just passing the IoU threshold. The delineation accuracy was also acceptable for the newly discovered rock glaciers in general, as shown in Figure 7c. However, we still conducted modifications to 100 out of the 123 landforms to ensure the quality of the mapping results after manual inspection (Figure 7d). The modification was made based on the Sentinel-2 optical images according to the geomorphic criteria presented in the IPA guideline (RGIK, 2021).