Scholarlink

<p>The increasing intensity and irregularity of precipitation driven by climate change have significantly elevated flash flood risk, particularly in semi-arid regions. This study aimed to model flash flood susceptibility (FFS) in the Konya Closed Basin, a semi-arid basin in Türkiye. A multi-layered dataset comprising topographic, geological, climatic, and environmental–anthropogenic conditioning factors was used to evaluate six ensemble machine learning models: Random Forest (RF), Extra Trees (ET), Rotation Forest (RotF), eXtreme Gradient Boosting (XGBoost), Light Gradient-Boosting Machine (LightGBM), and Hist Gradient Boosting (HistGB). Hyperparameter optimization was performed using GridSearchCV, and model performance was assessed under a five-fold spatial block cross-validation (SBCV) framework using classification metrics and ROC-AUC analysis. The results showed that the Extra Trees model achieved the highest AUC (0.851) and provided the most stable overall predictive performance under the SBCV-based evaluation framework. To improve the interpretability of model outputs, SHAP (Shapley Additive Explanations) was then applied to the best-performing ET model. SHAP analysis identified land use/land cover (LULC), lithology, and the Topographic Wetness Index (TWI) as the dominant controls on susceptibility. The findings indicate that FFS in the Konya Closed Basin is concentrated mainly in built-up areas and low-permeability lithological units where runoff generation and local wetness accumulation are enhanced, highlighting the combined roles of land-surface properties, infiltration limitation, flow concentration, and spatially structured validation in identifying flood-prone zones.</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

Flash flood susceptibility assessment in the semi-arid Konya Closed Basin (Türkiye) using ensemble machine learning and SHAP

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • İbrahim Çetin,
  • Emre Yılmaz,
  • Süleyman Sefa Bilgilioğlu

摘要

The increasing intensity and irregularity of precipitation driven by climate change have significantly elevated flash flood risk, particularly in semi-arid regions. This study aimed to model flash flood susceptibility (FFS) in the Konya Closed Basin, a semi-arid basin in Türkiye. A multi-layered dataset comprising topographic, geological, climatic, and environmental–anthropogenic conditioning factors was used to evaluate six ensemble machine learning models: Random Forest (RF), Extra Trees (ET), Rotation Forest (RotF), eXtreme Gradient Boosting (XGBoost), Light Gradient-Boosting Machine (LightGBM), and Hist Gradient Boosting (HistGB). Hyperparameter optimization was performed using GridSearchCV, and model performance was assessed under a five-fold spatial block cross-validation (SBCV) framework using classification metrics and ROC-AUC analysis. The results showed that the Extra Trees model achieved the highest AUC (0.851) and provided the most stable overall predictive performance under the SBCV-based evaluation framework. To improve the interpretability of model outputs, SHAP (Shapley Additive Explanations) was then applied to the best-performing ET model. SHAP analysis identified land use/land cover (LULC), lithology, and the Topographic Wetness Index (TWI) as the dominant controls on susceptibility. The findings indicate that FFS in the Konya Closed Basin is concentrated mainly in built-up areas and low-permeability lithological units where runoff generation and local wetness accumulation are enhanced, highlighting the combined roles of land-surface properties, infiltration limitation, flow concentration, and spatially structured validation in identifying flood-prone zones.