Scholarlink

<p>Mountain landslides are projected to become more frequent and severe due to climate change. However, few studies have applied geospatial artificial intelligence methods to generate landslide susceptibility maps (LSMs) under climate change scenarios. We addressed this gap by modeling landslide susceptibility in Kermanshah Province, Iran. A geodatabase of historical landslides was developed. Non-landslide points were generated using a hybrid method combining Voronoi mapping and Shannon entropy. Sixteen static factors (e.g., slope, geology, soil texture, distance to roads, drainage density) and two dynamic factors, annual precipitation and precipitation of the wettest month, were considered influential in landslide occurrence. For spatiotemporal modeling, the random forest (RF) algorithm was optimized using four novel metaheuristic algorithms, namely, ladybug beetle optimizer, electric eel feeding optimizer (EEFO), Harris hawk optimizer, and arithmetic optimizer algorithm (AOA). The performances of standard and optimized RF models were evaluated based on the area under the receiver operating characteristic curve (AUC-ROC), accuracy, precision, recall, specificity, and root mean square error (RMSE). Among the optimized models, RF-AOA and RF-EEFO performed best, achieving AUC-ROC values of 96.75% and 95.02%, respectively. While RF-EEFO demonstrated slightly higher accuracy and recall, with a lower RMSE than RF-AOA, the RF-AOA model had the highest value in precision and specificity during the testing phase. It indicates the superior performance of RF-AOA in accurately identifying landslide occurrences within high-susceptibility areas and reducing misclassification in low-susceptibility areas, thereby establishing it as the most reliable model for practical landslide susceptibility mapping. Therefore, for future susceptibility assessment, the RF-AOA model was selected, and LSMs were generated for 20-year intervals from 2020 to 2100 based on the HadGEM3-GC31-LL climate projection model. Results indicated that areas with high and very high susceptibility may increase from approximately 24% to 31% by 2100, whereas moderate susceptibility may rise from 20.60% to 24.80%. The feature importance analysis showed that hydro-topographic factors, particularly distance to drainages, annual precipitation, and slope, exert the strongest influence on the predictive ability of standard and optimized RF-based models. These findings provide valuable insights into the relationship between climate change and landslide-prone areas, supporting climate risk reduction, land use planning, and disaster management in mountainous regions.</p> Graphical Abstract <p></p> <p>The graphical abstract provides a comprehensive visual summary of an explainable geospatial artificial intelligence-based framework developed to model landslide susceptibility under climate change. Historical landslide inventories and 18 influential factors (both static, such as slope and geology, and dynamic, such as precipitation), were used as input data. The multicollinearity issue was examined using variance inflation factor (VIF) analysis to ensure model robustness. Landslide and non-landslide samples were split into training and testing subsets at a ratio of 80:20. A standard random forest (RF) model was developed and further optimized using four metaheuristic algorithms: arithmetic optimization algorithm (AOA), electric eel feeding optimizer (EEFO), Harris hawk optimizer (HHO), and ladybug beetle optimizer (LBO). The predictive performance of models was assessed using statistical metrics, including the area under the receiver operating characteristic curve (AUC-ROC) and root mean square error (RMSE). The RF-AOA and RF-EEFO models outperformed the others in detecting landslide-prone areas. Afterwards, future projections of landslide susceptibility maps were generated using RF-AOA method under three shared socioeconomic pathway (SSP) scenarios (SSP1-2.6, SSP2-4.5, SSP5-8.5) for four future time intervals (2021–2040, 2041–2060, 2061–2080, and 2081–2100). The results indicate an approximate 7% increase in areas with high and very high susceptibility by 2100. This research provides critical insights into climate-driven landslide hazards and offers a powerful decision-support tool for proactive land use planning and disaster risk mitigation in mountainous areas.</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

An Explainable Geospatial Artificial Intelligence Approach for Spatiotemporal Assessment of Climate Change Effects on Landslide Susceptibility

  • 期刊论文

  • 发表日期：

    2025年11月26日

查看全文
  • Zeynab Yousefi,
  • Ali Asghar Alesheikh,
  • Fatemeh Rezaie,
  • Saro Lee

摘要

Mountain landslides are projected to become more frequent and severe due to climate change. However, few studies have applied geospatial artificial intelligence methods to generate landslide susceptibility maps (LSMs) under climate change scenarios. We addressed this gap by modeling landslide susceptibility in Kermanshah Province, Iran. A geodatabase of historical landslides was developed. Non-landslide points were generated using a hybrid method combining Voronoi mapping and Shannon entropy. Sixteen static factors (e.g., slope, geology, soil texture, distance to roads, drainage density) and two dynamic factors, annual precipitation and precipitation of the wettest month, were considered influential in landslide occurrence. For spatiotemporal modeling, the random forest (RF) algorithm was optimized using four novel metaheuristic algorithms, namely, ladybug beetle optimizer, electric eel feeding optimizer (EEFO), Harris hawk optimizer, and arithmetic optimizer algorithm (AOA). The performances of standard and optimized RF models were evaluated based on the area under the receiver operating characteristic curve (AUC-ROC), accuracy, precision, recall, specificity, and root mean square error (RMSE). Among the optimized models, RF-AOA and RF-EEFO performed best, achieving AUC-ROC values of 96.75% and 95.02%, respectively. While RF-EEFO demonstrated slightly higher accuracy and recall, with a lower RMSE than RF-AOA, the RF-AOA model had the highest value in precision and specificity during the testing phase. It indicates the superior performance of RF-AOA in accurately identifying landslide occurrences within high-susceptibility areas and reducing misclassification in low-susceptibility areas, thereby establishing it as the most reliable model for practical landslide susceptibility mapping. Therefore, for future susceptibility assessment, the RF-AOA model was selected, and LSMs were generated for 20-year intervals from 2020 to 2100 based on the HadGEM3-GC31-LL climate projection model. Results indicated that areas with high and very high susceptibility may increase from approximately 24% to 31% by 2100, whereas moderate susceptibility may rise from 20.60% to 24.80%. The feature importance analysis showed that hydro-topographic factors, particularly distance to drainages, annual precipitation, and slope, exert the strongest influence on the predictive ability of standard and optimized RF-based models. These findings provide valuable insights into the relationship between climate change and landslide-prone areas, supporting climate risk reduction, land use planning, and disaster management in mountainous regions.

Graphical Abstract

The graphical abstract provides a comprehensive visual summary of an explainable geospatial artificial intelligence-based framework developed to model landslide susceptibility under climate change. Historical landslide inventories and 18 influential factors (both static, such as slope and geology, and dynamic, such as precipitation), were used as input data. The multicollinearity issue was examined using variance inflation factor (VIF) analysis to ensure model robustness. Landslide and non-landslide samples were split into training and testing subsets at a ratio of 80:20. A standard random forest (RF) model was developed and further optimized using four metaheuristic algorithms: arithmetic optimization algorithm (AOA), electric eel feeding optimizer (EEFO), Harris hawk optimizer (HHO), and ladybug beetle optimizer (LBO). The predictive performance of models was assessed using statistical metrics, including the area under the receiver operating characteristic curve (AUC-ROC) and root mean square error (RMSE). The RF-AOA and RF-EEFO models outperformed the others in detecting landslide-prone areas. Afterwards, future projections of landslide susceptibility maps were generated using RF-AOA method under three shared socioeconomic pathway (SSP) scenarios (SSP1-2.6, SSP2-4.5, SSP5-8.5) for four future time intervals (2021–2040, 2041–2060, 2061–2080, and 2081–2100). The results indicate an approximate 7% increase in areas with high and very high susceptibility by 2100. This research provides critical insights into climate-driven landslide hazards and offers a powerful decision-support tool for proactive land use planning and disaster risk mitigation in mountainous areas.