Texture-Based Region of Interest Extraction Using Unsupervised Clustering for Glacial Lake Detection

Abstract

Glacier melting has created glacial lakes which have become a significant issue of concern as far as the environment is concerned, mainly in high mountainous areas like the Himalayas, the Andes, and the Alps. The increased temperatures on the earth increase the melting of the glaciers and as a result glacial lakes are formed and grow. Such lakes can be unstable and lead to Glacial Lake Outburst Floods (GLOFs) which can cause drastic environmental and economic losses to the people down the stream. Thus, the timely detection and control of glacier lakes are required to manage risks of disasters and monitor the environment. The remote sensing technology is a viable tool of tracking glacial lakes in mountainous regions that are inaccessible and large. Nonetheless, it is difficult to detect glacial lakes correctly with the use of satellite images because of complex backgrounds made of snow, ice, shadows, rocks, and debris. In this paper, a new approach to automatic glacial lake detection is suggested based on multispectral satellite images and also parameterizing Region of Interest (ROI) with texture and using unsupervised clustering algorithms. In the suggested solution, the multispectral satellite images are initially subjected to preprocessing to improve the quality of images and eliminate noise. The features of texture are then extracted to obtain smooth surfaces areas that are most likely to be the water bodies. These areas of interest are termed as Regions of Interest (ROI) in an effort to minimize computational complexity. A clustering algorithm which has no supervision is then used to sort and categorize pixels as water and non-water with the help of spectral and texture data. The proposed method significantly outperforms traditional approaches such as NDWI, thresholding, Random Forest, and CNN models, achieving 98.45% accuracy, 99.14% precision, 97.70% recall, and 98.41% F1-score, ensuring highly reliable detection. It also demonstrates high computational efficiency, completing processing in just 29 seconds due to effective texture-based ROI reduction. Furthermore, the model shows strong generalization across multiple regions (Himalayas, Andes, Alps, Arctic) with very low error rates (4 false positives, 3 false negatives), confirming its robustness and accuracy in glacial lake detection. The presented system allows one to detect glacial lakes in an accurate way without using labeled training data. It has been proved through experimental analysis that the proposed method is more accurate than traditional spectral thresholding techniques in segmentation, and less prone to false detections. The method is especially applicable in the large scale observation of glacial lakes and can provide disaster mitigation measures in mountainous areas under threat.