Scholarlink

<p>Flash floods cause substantial economic losses and casualties worldwide. Modeling susceptibility to flash floods using hybrid models that combine statistical and machine learning methods is integral to flood mitigation strategies and disaster preparedness. Although the classification of flash flood conditioning factors is a critical step before applying such hybrid models, most previous studies have defaulted to using the natural breaks (NB) method for classification without exploring the potential impact of alternative discretization methods on model accuracy. Moreover, the classification system used to generate susceptibility maps determines the final appearance of the maps, which may influence decision-making tasks. In this context, this study introduces GeoDetector-based optimal discretization (OPGD) into factor classification and susceptibility partitioning. This approach is integrated with the Random Forest (RF) algorithm, resulting in the OPGD-RF model. Since the flexible number of classifications in OPGD-RF complicates objective comparisons, the GD-RF model was also constructed with a fixed number of classifications and compared with the NB-RF model built using the commonly used NB classification algorithm. The results show that GeoDetector-based models exhibit superior predictive performance in terms of area under the receiver operating characteristic curve (AUC = 0.946 and 0.934), followed by the NB-based model (AUC = 0.931). Additionally, OPGD-based partitioning results align well with historical flash flood events, explaining 64% of their spatial distribution. The application of GeoDetector provides a new perspective for improving the accuracy of integrated modeling and generating reliable susceptibility maps, in addition to providing effective support for rational resource allocation and targeted defense measures.</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

A flash flood susceptibility prediction and partitioning method based on GeoDetector and random forest

  • 期刊论文

  • 发表日期：

    2025年2月8日

查看全文
  • Xinyue Ke,
  • Ni Wang,
  • Yuan Xiu,
  • Zhenyu Dong,
  • Tingting Wang

摘要

Flash floods cause substantial economic losses and casualties worldwide. Modeling susceptibility to flash floods using hybrid models that combine statistical and machine learning methods is integral to flood mitigation strategies and disaster preparedness. Although the classification of flash flood conditioning factors is a critical step before applying such hybrid models, most previous studies have defaulted to using the natural breaks (NB) method for classification without exploring the potential impact of alternative discretization methods on model accuracy. Moreover, the classification system used to generate susceptibility maps determines the final appearance of the maps, which may influence decision-making tasks. In this context, this study introduces GeoDetector-based optimal discretization (OPGD) into factor classification and susceptibility partitioning. This approach is integrated with the Random Forest (RF) algorithm, resulting in the OPGD-RF model. Since the flexible number of classifications in OPGD-RF complicates objective comparisons, the GD-RF model was also constructed with a fixed number of classifications and compared with the NB-RF model built using the commonly used NB classification algorithm. The results show that GeoDetector-based models exhibit superior predictive performance in terms of area under the receiver operating characteristic curve (AUC = 0.946 and 0.934), followed by the NB-based model (AUC = 0.931). Additionally, OPGD-based partitioning results align well with historical flash flood events, explaining 64% of their spatial distribution. The application of GeoDetector provides a new perspective for improving the accuracy of integrated modeling and generating reliable susceptibility maps, in addition to providing effective support for rational resource allocation and targeted defense measures.