XGBoost-based method for flash flood risk assessment
XGBoost-based method for flash flood risk assessment
复制标题
DOI：
10.1016/j.jhydrol.2021.126382
复制
发表时间：
2021-05-17
影响因子：
6.4
通讯作者：
Wang, Zhongliang
中科院分区：
文献类型：
--
作者：
Ma, Meihong;Zhao, Gang;Wang, Zhongliang
山洪风险评估作为预防特大山洪灾害的一项广泛应用的技术，已成为当前的研究热点。然而，现有的评估山洪风险的机器学习方法大多依赖于单一分类器，适合处理小样本数据集，但所得的预测精度和泛化能力不足。与此同时，集成多个分类器的机器学习方法迄今为止还是未知的。极限梯度提升（XGBoost）是一种优秀的集成学习方法算法，在许多领域取得了令人瞩目的成果。它不仅优化了算法，还自动应用CPU的多线程进行并行计算，从而大大提高了模型训练速度和预测精度。因此，本文引入XGBoost模型进行山洪风险评估，然后结合两种输入策略和最小二乘支持向量机（LSSVM）模型验证其最优效果，从而提出基于XGBoost的山洪风险评估方法。随后，进行了归因分析，以评估这种方法可能出现的错误；最后，基于该方法生成了中国云南省县级山洪风险图。结果表明：（1）XGBoost表现良好，测试期间准确率为0.84，其五个指标（精确度、召回率、准确度、kappa和F-score）均高于LSSVM。 (2) 基于 XGBoost 的方法提供了可靠的山洪风险地图，并通过另一个山洪清单进行了验证，尽管一些错误可能归因于关键环境因素和统计灾害定位准确性。 （3）高风险县（含高风险县和最高风险县）占比40.3%，最高风险县主要集中在滇东南部。本文进一步解决了 XGBoost 的局限性（例如，作为一种耗时的贪心算法，不需要多线程优化）。以上结果表明，基于XGBoost的方法是获取高质量县级山洪风险图的有效方法，为我国正在进行的县级山洪防治工作提供了理论基础。
Flash flood risk assessment, a widely applied technology in preventing catastrophic flash flood disasters, has become the current research hotspot. However, most existing machine learning methods for assessing flash flood risk rely on a single classifier, which is suitable for processing small sets of sample data, but the resulting prediction accuracy and generalization ability are insufficient. Meanwhile, machine learning methods that integrate multiple classifiers are thus far unknown. Extreme Gradient Boosting (XGBoost) is an excellent algorithm for ensemble learning methods which has achieved remarkable results in many fields. It not only optimizes the algorithm but also automatically applies the CPU's multi-threading to perform parallel calculations, thus greatly improving the model training speed and prediction accuracy. Therefore, this article introduces the XGBoost model for the assessment of flash flood risk, and then combines the two input strategies and the Least squares support vector machine (LSSVM) model to verify its optimal effect, thus proposing the XGBoost-based method for flash flood risk assessment. Subsequently, an attribution analysis was implemented to assess the possible errors of this approach; and finally, a county-level flash flood risk map for Yunnan Province, China, was generated based on the proposed method. The results demonstrate that: (1) XGBoost performs well, with an accuracy of 0.84 in the testing period, and its five indices (precision, recall, accuracy, kappa, and F-score) are all higher than those of LSSVM. (2) The XGBoost-based approach provided the reliable flash flood risk maps, which were validated by another flash flood inventory, although some errors may be attributed to critical environmental factors and statistical disaster location accuracy. (3) The high-risk counties (including high-risk and highest-risk) accounted for 40.3%, with the highest-risk counties mainly concentrated in southeastern Yunnan. This article further addresses the limitations of XGBoost (e.g., as a time-consuming greedy algorithm, the non-necessity of multi-threaded optimization). All of the above results indicate that the XGBoost-based method is an effective method for obtaining high-quality county-level flash flood risk maps, which contributes to the theoretical basis for ongoing county-level flash flood prevention in China.