刘艳辉,方然可,苏永超,等. 2021. 基于机器学习的区域滑坡灾害预警模型研究[J]. 工程地质学报, 29(1): 116-124. doi: 10.13544/j.cnki.jeg.2020-533.
    引用本文: 刘艳辉,方然可,苏永超,等. 2021. 基于机器学习的区域滑坡灾害预警模型研究[J]. 工程地质学报, 29(1): 116-124. doi: 10.13544/j.cnki.jeg.2020-533.
    Liu Yanhui, Fang Ranke, Su Yongchao, et al. 2021. Machine learning based model for warning of regional landslide disasters[J]. Journal of Engineering Geology, 29(1): 116-124. doi: 10.13544/j.cnki.jeg.2020-533.
    Citation: Liu Yanhui, Fang Ranke, Su Yongchao, et al. 2021. Machine learning based model for warning of regional landslide disasters[J]. Journal of Engineering Geology, 29(1): 116-124. doi: 10.13544/j.cnki.jeg.2020-533.

    基于机器学习的区域滑坡灾害预警模型研究

    MACHINE LEARNING BASED MODEL FOR WARNING OF REGIONAL LANDSLIDE DISASTERS

    • 摘要: 中国滑坡灾害严重,区域滑坡灾害预警是防灾减灾的重要手段之一,预警模型是开展区域滑坡灾害预警的关键问题。本文系统开展了基于机器学习的区域滑坡灾害预警模型研究,并以四川省青川县为例,基于近10年地质与气象数据,构建了青川县区域滑坡灾害预警模型并开展实例校验。研究得出如下结论:(1)提出了基于机器学习的区域滑坡灾害预警模型的构建方法,主要包括训练样本集构建、样本学习训练与优化建模、模型保存与预警输出等几个关键步骤。(2)提出了区域滑坡训练样本集的构建方法,即以正样本为基础,在时空约束条件下随机采样获取负样本,最终获得完整的训练样本集。(3)样本学习训练中,以训练样本集的80%作为训练集,20%作为测试集,进行5折交叉验证,采用精确度、ROC曲线和AUC值校验模型准确度和模型泛化能力。采用贝叶斯优化算法进行模型优化。(4)实际预警中,调用训练好的预警模型输出滑坡灾害可能发生的概率。依据概率大小,分级确定预警等级。分级依据为:当输出概率P≥40%且P<60%时,发布黄色预警;当输出概率P≥60%且P<80%时,发布橙色预警;当输出概率P≥80%时,发布红色预警。(5)以青川县为例,构建了青川县区域滑坡训练样本集,采用6种机器学习算法进行模型训练,结果显示随机森林算法表现最好,其准确率最高(0.963),模型无过拟合现象,模型泛化能力最好(AUC=0.986);其次为逻辑回归算法;再次为人工神经网络算法和决策树算法。选取2018年6月26日的青川县日常预警业务进行实例校验。结果显示:当日17处滑坡灾害点中,100%的灾害点全部落入预警区范围内,其中:70.6%的滑坡落在红色预警区内,17.6%的滑坡落在橙色预警区内,11.8%的滑坡落在黄色预警区内。

       

      Abstract: The landslide disaster in China is serious, and early warning of regional landslide disaster is an important measure of disaster prevention and reduction. Then the study of the model is the key to carry out regional landslides warning successfully. This paper systematically proposes the construction method of the regional landslide disaster warning model based on machine learning, and takes Qingchuan County, Sichuan province as an example. Based on the geological and meteorological data of about ten years, the regional landslide disaster early warning model of Qingchuan County is built and an example verification is carried out. (1)The construction of regional landslide disaster warning model based on machine learning mainly includes several key steps including the construction of training sample set, sample training, parameter adjustment, model preservation and invocation. (2)A training sample set construction method for regional landslide early warning is proposed. That is, based on the positive samples, the negative samples are obtained by random sampling under spatial-temporal limitation, and the complete training sample set is finally obtained. (3)In the sample learning and training, 80% of the training sample set is taken as the training set, and 20% is taken as the test set, and 50-fold cross verification method is used. Then, Accuracy, ROC curve and AUC value verification model and model generalization ability are adopted. In order to achieve the optimal model, Bayesian Optimization Algorithm is used to optimize the model parameters. (4)In the early warning, the trained warning model is used to output the probability of landslide disaster. According to the probability, the early warning level is determined. When the output probability P≥40% and P < 60%, yellow alert can be issued; When the output probability P≥60% and P < 80%, an orange alert can be issued; and a red alert is issued when the output probability P≥80%. (5)Taking Qingchuan County as an example, six machine learning models were used for training. The results showed that the Random Forest model had the best performance, with the highest accuracy(0.963), no over-fitting phenomenon and the best generalization ability(AUC=0.986). The Second model is Logistic Regression model. Then, the model is artificial neural network model and decision tree model. We take the actual early-warning on June 26, 2018 as an example, load and call the pre-trained Random Forest model, calculate the output probability of the model, and divide the warning levels. According to the verification of the actual occurrence of landslide disasters, the result shows that all the landslides are within the warning area, among which 70.6% fall in the red warning area, 17.6% in the orange warning area and 11.8% in the yellow warning area.

       

    /

    返回文章
    返回