Implementation of stack-based ensemble technique for classification of glaciers in the western Himalayan catchments

Abstract

Human activities and climate change are causing Himalayan glaciers to melt erratically and affect runoff patterns, highlighting the need to monitor this vital resource. Imaging debris-covered glaciers is difficult because of the spectral similarity with non-glacier areas within various bands. This study assessed changes in the glacial area with regard to the year 1989 and mapped the area of glaciers in the Satluj River watershed from 2015 to 2019.The entire band range of Landsat imageries (1989–2019) was used to create glacial maps of each year, including glacier classes namely clean ice glaciers (CI), debris glaciers (DG), dirty ice + debris mix (DI + DM), glaciers and periglacial debris (PD), water, and rocks. The layers developed using traditional indices such as the NDSI, NDGI and unsupervised classification methods like K-means and Isodata. AI-powered technologies streamlined the process of mapping glacier borders and accurately assessed changes in glacial area. This work employs traditional machine learning techniques such as Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), Multi-layer Perceptron (MLP), and XGBoost (XGB), as well as a stack-based ensemble hybrid model. The six categorization systems' glacier class areas varied greatly, with accuracy ranging from 72.74% to 94.09%. The stack-based ensemble technique outperformed the other classification algorithms in this investigation. The transition from clean ice to dirty ice and eventually to a debris-covered glacier can also be observed in the basin. Overall, about 40–50% change (reduction) in the glacier area has been noticed.