ISSN 1000-3665 CN 11-2202/P
    董力豪,刘艳辉,黄俊宝,等. 基于卷积神经网络的福建省区域滑坡灾害预警模型[J]. 水文地质工程地质,2024,51(1): 145-153. DOI: 10.16030/j.cnki.issn.1000-3665.202211018
    引用本文: 董力豪,刘艳辉,黄俊宝,等. 基于卷积神经网络的福建省区域滑坡灾害预警模型[J]. 水文地质工程地质,2024,51(1): 145-153. DOI: 10.16030/j.cnki.issn.1000-3665.202211018
    DONG Lihao, LIU Yanhui, HUANG Junbao, et al. An early prediction model of regional landslide disasters in Fujian Province based on convolutional neural network[J]. Hydrogeology & Engineering Geology, 2024, 51(1): 145-153. DOI: 10.16030/j.cnki.issn.1000-3665.202211018
    Citation: DONG Lihao, LIU Yanhui, HUANG Junbao, et al. An early prediction model of regional landslide disasters in Fujian Province based on convolutional neural network[J]. Hydrogeology & Engineering Geology, 2024, 51(1): 145-153. DOI: 10.16030/j.cnki.issn.1000-3665.202211018

    基于卷积神经网络的福建省区域滑坡灾害预警模型

    An early prediction model of regional landslide disasters in Fujian Province based on convolutional neural network

    • 摘要: 福建省滑坡灾害频发,开展区域尺度上的滑坡灾害预警是防灾减灾的重要手段,但由于滑坡成灾机理复杂,传统的区域滑坡预警方法存在精度不足等问题。深度学习是指通过构建神经网络模型进行特征的提取、抽象、表示与学习的技术,是机器学习的一种。卷积神经网络作为一种经典的深度学习算法,具有比传统机器学习更强大的分类能力与表征能力。文章以福建省为研究区,将卷积神经网络引入滑坡灾害预警领域,构建福建省区域滑坡预警模型,过程及结果如下:(1)采用SMOTE优化算法对2010—2018年福建省滑坡灾害样本库进行优化,扩充正样本的个数,将正负样本比例从1∶3.4扩充到1∶2,样本总量达到18040个;(2)构建卷积神经网络模型结构,模型结构包括一个输入层、两个卷积层、两个最大池化层和一个全连接层以及一个输出层;(3)使用卷积神经网络对优化后的样本(2010—2018年样本的80%作为训练集)进行训练,并用贝叶斯优化算法优化模型超参数,得到福建省区域滑坡预警模型;(4)以2010—2018年样本的20%作为测试集对模型进行测试,采用混淆矩阵、ROC曲线进行模型测试,结果显示模型准确度为0.96~0.97,AUC值达到0.977,模型精度与泛化能力良好;(5)以2019年汛期滑坡灾害实况作为正样本,通过时空采样的方法采集负样本,构建2019年区域滑坡样本校验集(样本数603个),对模型进行进一步实况校验,采用混淆矩阵、ROC曲线进行模型校验,结果显示模型准确度为0.75~0.85,AUC值为0.852。虽然仅用了2019年汛期的滑坡实况样本进行校验,但也达到较好的效果。将卷积神经网络算法应用到区域滑坡预警中,为建立区域滑坡预警模型提供了一种新的途径,初步校验表明,模型效果良好,今后将在福建省对模型进行进一步的应用与校验。

       

      Abstract: Landslide disasters occur frequently in Fujian Province, and early warning of landslide disasters on a regional scale is an important means of effective disaster prevention and mitigation. Due to the complex mechanism of landslide disasters, the traditional regional landslide early warning methods have such problems as insufficient accuracy. Deep learning mainly refers to the technology of feature extraction, abstraction, representation and learning by constructing the neural network model, which is a kind of machine learning. As a classical deep learning algorithm, convolutional neural network has more powerful classification and representation ability than traditional machine learning. Taking Fujian Province as the research area, this paper introduces the convolution neural network into the field of landslide disaster early warning and constructs a regional landslide early warning model of Fujian Province. The process is as follows: (1) The SMOTE optimization algorithm is used to optimize the sample database of landslide disasters in Fujian Province from 2010 to 2018, enlarging the number of positive samples and expanding the proportion of positive and negative samples from 1∶3.4 to 1∶2, and the total number of samples reaches 18040. (2) Construct a convolution neural network model structure, which includes an input layer, two convolution layers, two maximum pooling layers, a full connection layer and an output layer. (3) Use the convolution neural network to train the optimized samples (80% of the samples from 2010 to 2018 as the training set), and use the Bayesian optimization algorithm to optimize the model parameters to obtain the regional landslide early warning model of Fujian Province. (4) The model is tested with 20% of the samples from 2010 to 2018 as the test set, and the confusion matrix and ROC curve are used to test the model. The results show that the accuracy of the model ranges from 0.96 to 0.97, the AUC value is 0.977, indicating that the model accuracy and generalization ability are good. (5) The actual situation of the landslide disaster in the flood season of 2019 is taken as a positive sample, negative samples are collected through the method of time-space sampling, and the 2019 regional landslide sample verification set (603 samples) is constructed. The model is further verified by using the confusion matrix and ROC curve. The results show that the accuracy of the model ranges from 0.75 to 0.85, and the AUC value is 0.852. Although only the actual landslide samples in the flood season of 2019 is used for verification, good results is also achieved. In this paper, the convolution neural network algorithm is applied to the regional landslide early warning, which provides a new way to establish the regional landslide early warning model. The preliminary verification shows that the model is effective and will be further applied and verified in Fujian Province in the future.

       

    /

    返回文章
    返回