Abstract
In the Karakoram Mountain range, glacial lakes are essential elements of the cryosphere. As a function of climate change and increasing temperature, these glacial lakes threaten downstream existence and the ecosystem by short time glacial lake outburst floods (GLOF). Therefore, the Glacial Lake mapping technique is a vital task to observe GLOF hazards. In this study, microwave Sentinel-1 Ground Range Detected (GRD) data used. It has the dual-polarization capability (HH + HV or VV + VH) and the ability to penetrate even through clouds or any weather condition. The study objective is to explore the application of GRD data and evaluate the efficiency and accuracy of machine learning algorithms for the extraction of water bodies. The study method is based on two main procedures, GRD backscattering analysis and supervised Machine Learning classifiers. The most commonly used machine learning classifiers are Random Forest (RF), K-nearest neighbor (KNN), and Maximum Likelihood. Although both procedures show better results for glacial lakes mapping in the study area, the mean backscatter parameter has the best accuracy rate than others in the total backscattering analysis. Likewise, in the classification approach, accuracy assessment was executed by comparing the results obtained for each classifier with the reference data. For all experiments, KNN performed the best at given training samples (Accuracy = 93%, Error rate = 0.06%) for both classes, compared to RF (Accuracy = 92%, Error rate = 0.07) and Maximum Likelihood (Accuracy = 90%, Error rate = 0.09%). The high classification accuracy obtained to extract glacial lakes using our approach will be useful to determine the short time flood outburst and take future precautionary measurements.