Scholarlink

<p>Human life, infrastructure, and environment face considerable damage in terms of rockfall. Conventional rockfall susceptibility mapping techniques, including statistics and deterministic approaches, cannot represent complex rockfall events spatial dependencies. To bridge this gap, this work integrates feature selection and optimization techniques with Graph Neural Networks (GNNs) for enhancing rockfall prediction accuracy. Four variants of GNNs, namely, Graph Convolutional Network (GCN), Graph Attention Network (GAT), GraphSAGE, and Graph Isomorphism Network (GIN), have been utilized for rockfall susceptibility prediction in Akhlamad Basin. Information Gain Ratio (IGR) and Symmetrical Uncertainty Attribute Evaluation (SUA) have been utilized for feature selection. Hyperparameter tuning using Optuna enhanced model performance. The results show that GraphSAGE outperformed other models, achieving the highest F1 Score (0.903), Kappa (0.800), and Accuracy (0.900). Key contributing factors to rockfall susceptibility included Topographic Position Index, frost, elevation, and distance to streams. The Friedman test and bootstrap analysis confirmed statistically significant differences in model performance, with GraphSAGE demonstrating superior generalization ability. Rockfall susceptibility maps derived from the models highlighted spatial variations in risk distribution, emphasizing the need for model specific thresholds in hazard assessment. This study demonstrates the efficacy of GNNs in geohazard prediction and provides a robust framework for improving rockfall susceptibility mapping.</p>

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

Optimizing graph neural networks for rockfall susceptibility mapping: a feature selection and hazard prediction approach

  • 期刊论文

  • 发表日期：

    2025年12月26日

查看全文
  • Leila Goli Mokhtari

摘要

Human life, infrastructure, and environment face considerable damage in terms of rockfall. Conventional rockfall susceptibility mapping techniques, including statistics and deterministic approaches, cannot represent complex rockfall events spatial dependencies. To bridge this gap, this work integrates feature selection and optimization techniques with Graph Neural Networks (GNNs) for enhancing rockfall prediction accuracy. Four variants of GNNs, namely, Graph Convolutional Network (GCN), Graph Attention Network (GAT), GraphSAGE, and Graph Isomorphism Network (GIN), have been utilized for rockfall susceptibility prediction in Akhlamad Basin. Information Gain Ratio (IGR) and Symmetrical Uncertainty Attribute Evaluation (SUA) have been utilized for feature selection. Hyperparameter tuning using Optuna enhanced model performance. The results show that GraphSAGE outperformed other models, achieving the highest F1 Score (0.903), Kappa (0.800), and Accuracy (0.900). Key contributing factors to rockfall susceptibility included Topographic Position Index, frost, elevation, and distance to streams. The Friedman test and bootstrap analysis confirmed statistically significant differences in model performance, with GraphSAGE demonstrating superior generalization ability. Rockfall susceptibility maps derived from the models highlighted spatial variations in risk distribution, emphasizing the need for model specific thresholds in hazard assessment. This study demonstrates the efficacy of GNNs in geohazard prediction and provides a robust framework for improving rockfall susceptibility mapping.