Scholarlink

<p>Landslides pose severe threats across the Tibetan Plateau, where rugged terrain, active tectonics, and fragile ecosystems jointly amplify hazard exposure. Although machine learning has been widely applied to landslide hazard assessment, limited attention has been given to how the structural properties of geospatial data influence model adaptability. In geospatial data processing, continuous topographic and environmental variables are typically discretized through gridding or sampling and reconstructed into structured tabular data with fixed fields and records. Building on this observation, this study introduces the concept of Geo-Tabular Data, characterized by low dimensionality, highly skewed feature distributions, and weak spatial continuity, and proposes a structure-aware framework to evaluate model adaptability. Using 30,084 labeled samples and 14 causative factors, we benchmark 14 models, spanning traditional machine-learning algorithms, ensemble tree-based methods, and deep tabular neural networks. The results show that GBDT models—LightGBM, XGBoost, and CatBoost—achieve the highest F1-scores (91.05%, 90.90%, and 90.31%, respectively), while deep tabular models such as TabNet, FT-Transformer, and TabTransformer perform lower (87.85%, 87.06%, and 80.14%). These discrepancies arise not from algorithmic complexity, but from structural misalignment between model architectures and the intrinsic properties of Geo-Tabular Data. By reframing landslide susceptibility modeling as a Geo-Tabular Learning task, this study underscores the importance of structure-driven model selection in Earth system science and provides practical guidance for large-scale geospatial risk mapping. The findings also highlight promising future directions, including temporal–dynamic integration, multimodal data fusion, and architecture design tailored to sparse and heterogeneous tabular environments.</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

Structure-driven evaluation of GBDT and neural networks for Geo-Tabular landslide hazard assessment on the Tibetan plateau

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Yaohui Liu,
  • Xinkai Wang,
  • Jie Zhou,
  • Huaqiao Xing,
  • Tongwen Liu,
  • Mengqiao He

摘要

Landslides pose severe threats across the Tibetan Plateau, where rugged terrain, active tectonics, and fragile ecosystems jointly amplify hazard exposure. Although machine learning has been widely applied to landslide hazard assessment, limited attention has been given to how the structural properties of geospatial data influence model adaptability. In geospatial data processing, continuous topographic and environmental variables are typically discretized through gridding or sampling and reconstructed into structured tabular data with fixed fields and records. Building on this observation, this study introduces the concept of Geo-Tabular Data, characterized by low dimensionality, highly skewed feature distributions, and weak spatial continuity, and proposes a structure-aware framework to evaluate model adaptability. Using 30,084 labeled samples and 14 causative factors, we benchmark 14 models, spanning traditional machine-learning algorithms, ensemble tree-based methods, and deep tabular neural networks. The results show that GBDT models—LightGBM, XGBoost, and CatBoost—achieve the highest F1-scores (91.05%, 90.90%, and 90.31%, respectively), while deep tabular models such as TabNet, FT-Transformer, and TabTransformer perform lower (87.85%, 87.06%, and 80.14%). These discrepancies arise not from algorithmic complexity, but from structural misalignment between model architectures and the intrinsic properties of Geo-Tabular Data. By reframing landslide susceptibility modeling as a Geo-Tabular Learning task, this study underscores the importance of structure-driven model selection in Earth system science and provides practical guidance for large-scale geospatial risk mapping. The findings also highlight promising future directions, including temporal–dynamic integration, multimodal data fusion, and architecture design tailored to sparse and heterogeneous tabular environments.