Scholarlink

<p>Debris flows pose persistent threats to transport corridors, settlements, and engineering facilities in the northern sector of the China-Pakistan Economic Corridor, where extreme relief, active tectonics, dense drainage, and strong hydroclimatic forcing create strong spatial heterogeneity in environmental predisposition. In this setting, the central challenge is not simply to produce another susceptibility map, but to establish a defensible workflow for inventory-based regional assessment and management-oriented screening. This study develops DOME, the Dynamic Optimization Meta-Ensemble framework, within a rigorously separated development and evaluation design. Using 3447 documented debris-flow units, 3447 pseudo-absence units, and 11 environmental conditioning factors, DOME integrates dynamic learner screening, multi-objective ensemble optimization, leak-free out-of-fold stacking, isotonic calibration, and multiple validation protocols to generate calibrated relative susceptibility scores. Four socioeconomic variables are excluded from model training and introduced only in a subsequent exposure integration stage. On the independent test set, DOME achieved a ROC AUC of 0.9828, a PR AUC of 0.9794, an F1 score of 0.9328, an MCC of 0.8656, a Brier score of 0.0498, and a log loss of 0.1701. Under blocked spatial cross-validation, mean ROC AUC and PR AUC remained 0.9662 and 0.9587, demonstrating strong within-study-area spatial transferability. Regional deployment revealed a corridor controlled susceptibility pattern, with elevated scores concentrated along dissected valley systems, tributary corridors, and structurally complex mountain fronts. SHAP analysis showed that basin area, main channel bed gradient, elevation range, fault density, and lithology dominate the fitted susceptibility function through pronounced nonlinear responses and pairwise interaction structures. The calibrated susceptibility surface was then integrated with an independently constructed exposure proxy surface to derive a management oriented prioritization layer for corridor scale screening. The study provides a robust, interpretable, and operational framework for regional debris flow susceptibility assessment in high mountain corridor systems.</p> Graphical abstract <p></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

DOME: a dynamic optimization meta-ensemble framework for debris-flow susceptibility assessment with SHAP-based interpretation

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Tianlong Wang,
  • Hao Yang

摘要

Debris flows pose persistent threats to transport corridors, settlements, and engineering facilities in the northern sector of the China-Pakistan Economic Corridor, where extreme relief, active tectonics, dense drainage, and strong hydroclimatic forcing create strong spatial heterogeneity in environmental predisposition. In this setting, the central challenge is not simply to produce another susceptibility map, but to establish a defensible workflow for inventory-based regional assessment and management-oriented screening. This study develops DOME, the Dynamic Optimization Meta-Ensemble framework, within a rigorously separated development and evaluation design. Using 3447 documented debris-flow units, 3447 pseudo-absence units, and 11 environmental conditioning factors, DOME integrates dynamic learner screening, multi-objective ensemble optimization, leak-free out-of-fold stacking, isotonic calibration, and multiple validation protocols to generate calibrated relative susceptibility scores. Four socioeconomic variables are excluded from model training and introduced only in a subsequent exposure integration stage. On the independent test set, DOME achieved a ROC AUC of 0.9828, a PR AUC of 0.9794, an F1 score of 0.9328, an MCC of 0.8656, a Brier score of 0.0498, and a log loss of 0.1701. Under blocked spatial cross-validation, mean ROC AUC and PR AUC remained 0.9662 and 0.9587, demonstrating strong within-study-area spatial transferability. Regional deployment revealed a corridor controlled susceptibility pattern, with elevated scores concentrated along dissected valley systems, tributary corridors, and structurally complex mountain fronts. SHAP analysis showed that basin area, main channel bed gradient, elevation range, fault density, and lithology dominate the fitted susceptibility function through pronounced nonlinear responses and pairwise interaction structures. The calibrated susceptibility surface was then integrated with an independently constructed exposure proxy surface to derive a management oriented prioritization layer for corridor scale screening. The study provides a robust, interpretable, and operational framework for regional debris flow susceptibility assessment in high mountain corridor systems.

Graphical abstract