Systematic Extraction and Outburst Susceptibility Assessment of Glacier Lakes using Satellite Images

Abstract

The glacial lakes on the Himalayan region are increasing continuously. Various factors are responsible for increasing the glacial lakes including glacial retreat, glacia lake outburst floods (GLOFs), increased rainfall and human activities. The expansion of glacial lakes in the Himalayas has raised concerns about potential GLOFs and their impacts on downstream communities, infrastructure, and ecosystems. Monitoring and studying these glacial lakes is crucial for understanding the changing dynamics of the Himalayan region and for developing strategies to mitigate the risks associated with glacial lake growth. Timely and accurate extraction of glacial lake information is cru- cial for understanding the complex interactions between natural phenomena and human activities in the Himalayan region. However, monitoring glacial lakes through tradi- tional field surveys can be time-consuming, costly, and challenging, especially when dealing with vast or remote areas in the Himalayan region. Indeed, advanced remote sensing technologies have revolutionized ability to monitor and understand earth’s sur- face dynamics effectively. The regular acquisition of remote sensing images over the same geographical regions offers a wealth of valuable information that can be applied across various fields, including environmental science, geology, agriculture, urban plan- ning, disaster management, and more. The vast amounts of data generated by remote sensing technologies and the need for timely and accurate monitoring of land surface transitions, the development of automatic techniques is essential. Manual monitoring of remote sensing images on a large scale is not only time-consuming but also prone to human error. Therefore, development of automatic techniques are essential, which can extract and monitor the glacial lake region with minimal human involvement. However, extracting glacial lakes from multitemporal remote sensing images involves specialized techniques to identify and delineate these features accurately. These techniques use the differences in spectral and spatial characteristics between glacial lakes and their surroundings. This thesis is focused on extracting the regions of glacial lakes from multispectral remote sensing images, which is a significant and valuable research topic, especially in the context of monitoring and understanding glacial lake dynamics in the Himalayan region and other glaciated areas. To achieve accurate extraction of glacial lake regions from satellite imagery, this study first introduces a modified normalized cut (Ncut) segmentation technique that integrates region adjacency graph (RAG) and simple linear iterative clustering (SLIC). While traditional Ncut methods often exhibit weak adherence to image boundaries, the proposed approach enhances segmentation precision by globally optimizing a cost function that evaluates intra-group similarity and inter-group dissimilarity. In this enhanced framework, SLIC efficiently generates superpixels through k-means clustering, reducing computational complexity. These su- perpixels are then grouped using RAG, which merges adjacent regions based on SLIC labels, thereby simplifying the graph structure. The Ncut algorithm is applied not to individual pixels but to these merged regions, effectively mitigating common issues like pixel misclassification and background heterogeneity in satellite images. Once the glacial lakes are accurately delineated, a threshold-based method is used to estimate their out- burst susceptibility index. Building upon this segmentation groundwork, the research further explores advanced data-driven techniques to improve performance. The previous models relied on rule-based methods and manual parameter tuning; to overcome this, a deep learning-based approach that surpasses such limitations is proposed. Specifi- cally, the glacial lakes U-Net (GLU-Net) architecture is introduced, which uses semantic learning to enhance glacial lake segmentation. Unlike earlier rule-dependent methods, GLU-Net employs skip connections to retain crucial spatial features that are typically lost during max-pooling operations. This results in a more accurate and robust extrac- tion process, significantly reducing the reliance on manual intervention and parameter adjustment. The vast dataset is prepared with the help of various water indices for proper training in this study. In this work, different deep learning (DL) segmentation architectures with various encoders are demonstrated. The qualitative and quantitative performance evaluations reveal that the proposed system achieves effective extraction of glacial lake regions without relying on manual field surveys. This autonomous ca- pability marks a significant advancement in remote sensing-based analysis. To further evaluate the potential risk associated with these lakes, a support vector machine (SVM) classifier is introduced to compute the GLOF susceptibility index. Although GLU-Net exhibits strong segmentation accuracy, its lack of attention mechanisms limits its ability to fully capture both global and local spatial features critical for detailed glacial lake de- lineation and risk assessment. To address this limitation, the study advances to a more sophisticated architecture global-local salient mapping (GLSM-Net) is proposed, which incorporates a hybrid deep learning framework that uses both global and local feature representations. This model introduces a global feature salient map (GFSM) and a lo-cal feature salient map (LFSM), designed to jointly extract high-level contextual cues and fine-grained spatial details. The architecture builds upon an enhanced U-Net++ backbone, utilizing SE-Net modules locally and integrating feature cross-fusion blocks (FCFB) to effectively merge multi-scale spatial information. This dual-path approach significantly improves the accuracy of lake boundary detection. Furthermore, the model computes GLOF risk scores through dense layers following feature extraction, enabling precise flood risk estimation. The experimental results confirm the robustness of GLSM-Net in both glacial lake segmentation and flood risk prediction, offering valuable tools for disaster preparedness in high-altitude environments. GLSM-Net emphasizes the uti-lization of attention-based spatial features for assessing the risk of GLOFs, effectively enhancing the identification of vulnerable areas based on spatial cues. However, its limitation lies in the exclusion of historical data, which restricts its ability to capture temporal variations and evolving risk patterns over time. Addressing this gap, the proposed work introduces adaptable dilated U-Net (AdU-Net), an advanced framework that combines a dilated U-Net with nested skip connections and an adaptable vision transformer encoder (AVi-TE) to improve glacial lake extraction from satellite imagery. By incorporating transformer-based encoding, AdU-Net captures both long-range de-pendencies and fine-grained spatial hierarchies, yielding high-precision segmentation of glacial lake boundaries. To complement spatial analysis with temporal insight, a modi-fied spiking neural network (SNN) is further proposed for GLOF risk evaluation. SNNs are well-suited for processing temporally sequenced data and identifying dynamic pat-terns, enabling a more holistic understanding of outburst susceptibility. The integrated system is applied to Landsat 8 satellite imagery across the Imja, Chandra, and Bhaga glacial regions, offering a comprehensive approach that merges spatial attention, tempo-ral dynamics, and adaptive modeling to advance GLOF risk assessment. Experimental evaluations on multispectral data validate the effectiveness of this hybrid methodology, highlighting its potential for informed flood risk mitigation strategies in glaciated high- altitude zones.