Integration of optical, SAR and DEM data for automated detection of debris-covered glaciers over the western Nyainqentanglha using a random forest classifier

Abstract

The capability of optical images to detect clean-ice glaciers has been well-demonstrated. Debris-covered glaciers, are different from clean-ice glaciers and are important for research of glacier mass balance. However, the main challenges to detect debris-covered glaciers in alpine regions faced by many researchers are the spectral similarity between debris-covered glaciers and the rocks and soils on both sides, as well as shadows cast from mountains, clouds and seasonal snow in satellite images. This study aimed to develop an automatic algorithm (using a Random Forest classifier model) implemented in the western Nyainqentanglha to map debris-covered glaciers based on multi-source datasets such as Sentinel-1 Synthetic Aperture Radar data, Sentinel-2 Multispectral Instrument data, Landsat-8 Thermal Infrared Sensor and Digital Elevation Models. All data was split into training (70%) and testing datasets (30%). The main strength of this study is that our method overcomes most of the above-mentioned challenges and the great accuracy (Kappa coefficient: 0.96, overall accuracy: 97.21%) of the Random Forest classifier model represents a comprehensive success in identifying debris-covered glaciers, illustrating that if this method can be executed efficiently, it will bring benefits for glacier inventory management. Additionally, an analysis of the spatial characteristic of the mountain glaciers showed that the glacier area, elevation and slope of the Nyainqentanglha Mountains were closely related. Moreover, most glacier movement occurred in glaciers with an area of less than 1 km2 and greater than 10 km2, which had mean velocities of 0.141 m/day and 0.168 m/day, respectively, and glacier movement explained the uncertainty of glacier facies mapping.