高级检索

    统计模型与机器学习模型在泥石流易发性评价中的耦合效应研究

    Coupling effect of statistical models and machine learning models in the evaluation of debris flow susceptibility

    • 摘要:
      研究目的 怒江流域云南段位于西部横断山脉,以深切高山峡谷地貌为主,一直是泥石流的高发区,开展区内泥石流易发性精准评价对区内防灾减灾具有重要意义。
      研究方法 本文基于地质调查工程获取数据,以流域为评价单元,利用信息量模型、频率比模型及确定性系数模型等统计模型与随机森林模型、BP神经网络模型等机器学习模型耦合,分别对怒江干流流域内泥石流灾害易发性进行评价,通过模型评估参数、种子面积单元指数及ROC曲线对各类耦合模型评价精度及可靠性进行了对比分析,进而确定最优耦合集成模型。
      研究结果 各类集成模型评价效果优越性顺序表现为频率比−随机森林>信息量−随机森林>确定性系数−随机森林>频率比−神经网络>信息量−神经网络>确定性系数−神经网络。
      结论 频率比−随机森林模型在所有集成模型中预测精度及评价效果最优;在同一统计模型条件下,随机森林模型相较于神经网络模型具有明显的优越性;在统计模型处理后数据与机器学习模型的适配性上,频率比模型优于信息量模型优于确定性系数模型。

       

      Abstract:
      This paper is the result of geological hazard survey engineering.
      Objective The Yunnan section of the Nujiang River Basin is located in the western Hengduan Mountains. It is dominated by deep mountain canyon landforms and has always been a high-incidence area of debris flows, so it is of great significance to carry out an accurate evaluation of mudslide susceptibility in the region for hazard prevention and mitigation.
      Methods Based on the data obtained from geological survey projects, this study takes the basin as the evaluation unit, and uses statistical models such as information volume model, frequency ratio model and certainty coefficient model to couple with machine learning models such as random forest model and BP neural network model to evaluate the susceptibility of debris flow hazards in the mainstream Nujiang River Basin. The evaluation accuracy and reliability of various coupling models are compared and analyzed through model evaluation parameters, seed area unit index and ROC curve before we have identified the optimal coupling integrated model(s).
      Results The ranking of various integrated models from best to worst in terms of evaluation effect is as follows: Frequency ratio-Random forest > Information volume-Random forest > Certainty coefficient-Random forest > Frequency ratio-Neural network > Information volume-Neural network>Certainty coefficient-Neural network.
      Conclusions The frequency ratio-random forest model has the best prediction accuracy and evaluation efficacy among all integrated models; Under the same statistical model conditions, the random forest model demonstrates a significant advantage over the neural network model; And furthermore, in terms of the adaptability of data after statistical model processing and machine learning model, the frequency ratio model performs better than the information volume model, which in turn performs better than the certainty coefficient model.

       

    /

    返回文章
    返回