Spatio-Temporal Estimation of Glacier Dynamics under Climate Change Scenarios Using Machine Learning Techniques

Glaciers of the Upper Indus Basin (UIB) play a vital role in providing water resources, hydropower generation, and livelihood, but they are very vulnerable and sensitive to continuous climate change impacts. This research presents a novel approach for accurate mapping of glacier extent, clean ice, debris cover, seasonal snow, and glacier melt across the Hunza Basin. We have used Grey Level Co-occurrence Matrix (GLCM), Machine Learning (ML) techniques of Random Forest (RF), Artificial Neural Networks (ANN), and Support Vector Machines (SVM) to conduct the purposeful research. ML models were trained on multispectral (Landsat, Sentinel-1 & 2, MODIS, and SPOT-5 from the last 35 years) and textural datasets. Overall, 6628 samples for training and 988 samples for testing were used to maintain a 70/ 30 ratio to evaluate overall accuracy (OA) and kappa coefficient (k̂). RF ensured the best results (OA = 95.4 %, k̂ = 0.965) in comparison of ANN (OA = 94%, k̂ = 0.92) and SVM (OA = 92 %, k̂ = 0.89). The accuracy of clean ice and seasonal snow remained consistent (producer accuracy and user accuracy >93%) compared to that of debris cover and glacier melt. Glacier retreat, increased ablation, formation of clean ice loss, and frequency of supraglacial melt due to expansion of debris cover up to 23.31% were witnessed spatially in the basin. Proposed approaches prove that ML techniques are very useful for the estimation of risk assessment in the climate-prone mountain basins and offer a robust way forward for hydrological modelling, glacier change monitoring, and water resource management.