Scholarlink

This paper introduces an innovative deep learning pipeline designed to predict glacier melting in the Himalayan region using satellite imagery and advanced neural network architectures. Focusing on the Gangotri glacier, we extract over 200 sequential frames from Sentinel-based timelapse GIFs captured between 2020 and 2025. Since ground-truth melt values are unavailable, we simulate melt progression by calculating normalized pixel-wise differences between consecutive frames. To extract high-level spatial features, we utilize a pretrained ResNet50 Convolutional Neural Network (CNN). These features are then fed into a Spatio-Temporal Graph Neural Network (ST-GNN) built using the GraphSAGE framework, allowing the model to learn both spatial and temporal dependencies across glacier image sequences. Our proposed ST-GNN model achieves a melt prediction accuracy of 79%, demonstrating strong generalization and performance in handling real satellite data. For comparative analysis, a CNN+LSTM-based baseline model is also implemented, which attains a lower accuracy of 59%. The results clearly highlight the superior ability of the ST-GNN architecture in capturing complex glacier melt patterns over time, making it a promising approach for climate change monitoring and forecasting.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting Himalayan Glacier Melting with Spatio-Temporal Graph Neural Networks Using Satellite Imagery

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pasupuleti Hari Samhita,
  • R. Rathna

摘要

This paper introduces an innovative deep learning pipeline designed to predict glacier melting in the Himalayan region using satellite imagery and advanced neural network architectures. Focusing on the Gangotri glacier, we extract over 200 sequential frames from Sentinel-based timelapse GIFs captured between 2020 and 2025. Since ground-truth melt values are unavailable, we simulate melt progression by calculating normalized pixel-wise differences between consecutive frames. To extract high-level spatial features, we utilize a pretrained ResNet50 Convolutional Neural Network (CNN). These features are then fed into a Spatio-Temporal Graph Neural Network (ST-GNN) built using the GraphSAGE framework, allowing the model to learn both spatial and temporal dependencies across glacier image sequences. Our proposed ST-GNN model achieves a melt prediction accuracy of 79%, demonstrating strong generalization and performance in handling real satellite data. For comparative analysis, a CNN+LSTM-based baseline model is also implemented, which attains a lower accuracy of 59%. The results clearly highlight the superior ability of the ST-GNN architecture in capturing complex glacier melt patterns over time, making it a promising approach for climate change monitoring and forecasting.