Scholarlink

<p>Landslide susceptibility mapping in mountainous transportation corridors requires modelling approaches that capture complex nonlinear relationships and spatial dependencies among geo-environmental factors. This study proposes a hybrid graph-based artificial intelligence framework (GraphSAGE–CatBoost optimized by the Reptile Search Algorithm) for landslide susceptibility assessment along the Karaj–Chalus road, a critical transportation corridor in the central Alborz mountains, northern Iran. The framework explicitly models terrain units as nodes in a spatially structured graph, enabling the capture of topological dependencies that are inherently inaccessible to conventional pixel-based or tabular approaches. A comprehensive database comprising of 409 documented landslides, 409 non-landslide samples, and ten conditioning factors (lithology, distances to faults, roads and rivers, land use/land cover, rainfall, slope, aspect, topographic wetness index, and stream power index) was compiled for model development. To rigorously benchmark the proposed graph-based model against complementary state-of-the-art paradigms, two representative models were selected: (1) a GoogleNet-CNN optimized by Harris Hawks Optimization, representing deep convolutional approaches with multi-scale feature extraction; and (2) an Autoencoder–XGBoost hybrid, representing hybrid deep-ensemble strategies with unsupervised feature compression. Under random split validation (70/30), the GraphSAGE–CatBoost–RSA model achieved superior performance across multiple metrics: AUC-ROC = 0.972 (95% bootstrap CI [0.963–0.981]), AUC-PR = 0.962 ([0.951–0.973]), accuracy = 0.932, precision = 0.915, recall = 0.948, F1 = 0.931, and test error = 0.068, outperforming both benchmarks with non-overlapping confidence intervals. Importantly, spatial cross-validation using three geographically independent sub-regions confirmed that these results are not due to spatial leakage; the model maintained strong generalization (mean AUC-ROC = 0.924, mean accuracy = 0.886, mean F1 = 0.885), with only a 5% reduction from the random-split performance. Geomorphologically, high-susceptibility zones are strongly associated with steep slopes underlain by weak clay-rich lithologies, proximity to road cuts and drainage networks, and north-facing aspects where soil moisture persists. The resulting susceptibility map identifies priority segments for targeted mitigation, including tunnel portals, bridge abutments, and cut-slopes requiring drainage upgrades or stabilization. The proposed framework provides a spatially explicit and rigorously validated modelling approach for landslide susceptibility mapping in mountainous transportation corridors, offering a robust basis for comparing graph-based learning with established deep-learning and hybrid machine-learning alternatives.</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

Advanced graph-based and deep learning frameworks for landslide susceptibility mapping in mountain transportation corridors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月2日

查看全文
  • Yousef Bahrami,
  • Abbas Maghsoudi,
  • Amin Beiranvand Pour,
  • Masoud Shirali

摘要

Landslide susceptibility mapping in mountainous transportation corridors requires modelling approaches that capture complex nonlinear relationships and spatial dependencies among geo-environmental factors. This study proposes a hybrid graph-based artificial intelligence framework (GraphSAGE–CatBoost optimized by the Reptile Search Algorithm) for landslide susceptibility assessment along the Karaj–Chalus road, a critical transportation corridor in the central Alborz mountains, northern Iran. The framework explicitly models terrain units as nodes in a spatially structured graph, enabling the capture of topological dependencies that are inherently inaccessible to conventional pixel-based or tabular approaches. A comprehensive database comprising of 409 documented landslides, 409 non-landslide samples, and ten conditioning factors (lithology, distances to faults, roads and rivers, land use/land cover, rainfall, slope, aspect, topographic wetness index, and stream power index) was compiled for model development. To rigorously benchmark the proposed graph-based model against complementary state-of-the-art paradigms, two representative models were selected: (1) a GoogleNet-CNN optimized by Harris Hawks Optimization, representing deep convolutional approaches with multi-scale feature extraction; and (2) an Autoencoder–XGBoost hybrid, representing hybrid deep-ensemble strategies with unsupervised feature compression. Under random split validation (70/30), the GraphSAGE–CatBoost–RSA model achieved superior performance across multiple metrics: AUC-ROC = 0.972 (95% bootstrap CI [0.963–0.981]), AUC-PR = 0.962 ([0.951–0.973]), accuracy = 0.932, precision = 0.915, recall = 0.948, F1 = 0.931, and test error = 0.068, outperforming both benchmarks with non-overlapping confidence intervals. Importantly, spatial cross-validation using three geographically independent sub-regions confirmed that these results are not due to spatial leakage; the model maintained strong generalization (mean AUC-ROC = 0.924, mean accuracy = 0.886, mean F1 = 0.885), with only a 5% reduction from the random-split performance. Geomorphologically, high-susceptibility zones are strongly associated with steep slopes underlain by weak clay-rich lithologies, proximity to road cuts and drainage networks, and north-facing aspects where soil moisture persists. The resulting susceptibility map identifies priority segments for targeted mitigation, including tunnel portals, bridge abutments, and cut-slopes requiring drainage upgrades or stabilization. The proposed framework provides a spatially explicit and rigorously validated modelling approach for landslide susceptibility mapping in mountainous transportation corridors, offering a robust basis for comparing graph-based learning with established deep-learning and hybrid machine-learning alternatives.