Automated Glacial Lake Mapping in the Himalayas: An Ensemble Multi-Sensor Approach with Random Forest and High-Resolution PlanetScope Imagery Bhawna Pathak , Ankit Singh, Dericks P. Shukla
Dexter Lab, School of Civil and Environmental Engineering, Indian Institute of Technology (IIT) Mandi, 175005, Himachal, India.
Abstract
Glacial lakes are critical indicators of climate change and present significant Glacial Lake Outburst Flood (GLOF) risks in high-mountain areas. Accurate and automated mapping and monitoring of these dynamic features is crucial but challenged by complex terrain, persistent cloud cover, and spectral ambiguities. This study introduces an automated method for detecting glacial lakes in Northwestern Himalaya. This method leverages multi-source remote sensing data and a robust Random Forest (RF) classifier.
Our approach introduces a classification method that integrates an ensemble of Sentinel-1 Synthetic Aperture Radar (SAR), Sentinel-2 Multi-spectral Instrument (MSI), SRTM Digital Elevation Model (DEM), and high-resolution 3-meter PlanetScope optical imagery. This data fusion offers exceptional detail necessary for accurately defining glacial lake boundaries.
The RF model, trained on an augmented dataset, tackles the problem of misclassifications involving streams and wet surfaces- it demonstrates impressive results. The model achieved an overall accuracy of 94.44%, along with precision, recall, and F-1 scores of 0.95, 0.97, and 0.96, respectively, and an AUC-ROC score of 0.983.
This method demonstrates clear advantages over existing approaches. Deep learning models like GLNet require large amounts of labeled training data, high computational resources, and specialized GPU infrastructure. On the other hand, our machine learning model (RF) offers comparable performance without such intensive requirements, making it more accessible and practical for broader glaciological applications. It effectively deals with common issues such as misclassifications of shadows, supraglacial melt ponds, and streams. Furthermore, the approach is temporally transferable and can be adapted for multi-temporal analysis of glacial lake dynamics. This offers valuable insights into long-term monitoring and integration into early warning systems and GLOF risk reduction frameworks. This scalable and interpretable workflow provides a practical alternative to deep learning models, supporting high-resolution glacial lake inventories.
Keywords: Glacial lakes, GLOF, automated detection, Random Forest.
Topic: Topic C: Emerging Technologies in Remote Sensing
