基于OOD泛化性验证和深度全连接神经网络的泥石流易发性评价方法  

Methods for Evaluating Debris Flow Susceptibility Based on OOD Generalization Verification and Deep Fully Connected Neural Networks

作　　者：郭鹏宁 邢会歌[1] 李从江 吴雨鑫 李海波 GUO Pengning;XING Huige;LI Congjiang;WU Yuxin;LI Haibo(College of Architecture and Environment,Sichuan Univ.,Chengdu 610065,China;College of Water Resource&Hydropower,Sichuan Univ.,Chengdu 610065,China;State Key Lab.of Hydraulics and Mountain River Eng.,Sichuan Univ.,Chengdu 610065,China)

机构地区：[1]四川大学建筑与环境学院,四川成都610065 [2]四川大学水利水电学院,四川成都610065 [3]四川大学山区河流保护与治理全国重点实验室,四川成都610065

出　　处：《工程科学与技术》2024年第4期182-193,共12页Advanced Engineering Sciences

基　　金：国家自然科学基金区域创新发展联合基金(U20A20111);四川省青年科技创新研究团队项目(2020JDTD0006)。

摘　　要：提升易发性评价精度有助于山区泥石流灾害早期的识别和监测预警。大部分机器学习模型在训练、测试集合上表现良好,但实际应用过程精度较差,不利于工程选址规划和防灾减灾,如何提高机器学习模型评价精度与泛化性具有重要意义。选取深度全连接神经网络,与梯度提升树、随机森林模型和贝叶斯网络等机器学习方法共同进行模型精确性评价和OOD(out-of-distribution)泛化性验证,从而找出在训练、预测和应用中均具有较高精度的方法。以四川省雅安市为例,采用小流域单元进行区域网格划分,将数据集合按7∶3比例随机分为训练集和测试集,使用经验法则(3-sigma)剔除异常数据,并基于多变量(Iterative Imputer)和K-近邻法对缺失值填充进行泥石流灾害易发性评价。在泥石流易发性因子的共线性、敏感性和预测能力的分析结果基础上,选定14个易发性因子构建模型评价指标体系,进行泥石流易发性评价与对比。通过对模型的精确性评价及OOD泛化性验证发现:深度全连接神经网络模型曲线下的面积(AUC)、准确率(Acc)、召回率(Recall)的值比梯度提升树等的计算结果分别超出了0.027、0.02、0.02,而平均绝对值误差(MAE)降低了0.003;OOD泛化性验证准确度超出了0.056。研究表明,深度全连接神经网络对于泥石流易发性评价的预测效果较好,能够提高泥石流评价的精度,增加评价的适应性,可为泥石流易发性评价提供新思路。More-accurate predictions of susceptibility to debris flows would help greatly in the early identification and large-scale monitoring of debris-flow disasters.Some existing models for doing so perform well during training and test but less well in practice,and this has adverse effects on engineering site selection and disaster prevention and mitigation.To find a method with high accuracy in training,prediction,and application,the present study considered four models—i.e.,deep fully connected neural network,gradient boosting decision tree,random forest,and Bayesian network and assessed them for accuracy and out-of-distribution(OOD)generalization.Taking Ya’an City in China’s Sichuan Province as an example,small watershed units were used for regional meshing,and the data were divided randomly into a training set and a test set at the ratio of 7:3.During data cleansing,missing values were tackled using a K-means and IterativeImputer method,and excrescent data were rejected using the 3-sigma rule.After analyzing the collinearity,sensitivity,and predictive ability of various debris-flow susceptibility factors,14 were selected for model operation.The four aforementioned models were then constructed for debris-flow susceptibility prediction and model comparison,and accuracy evaluation and OOD generalization verification showed that the deep fully connected neural network model was superior to the other three models in terms of AUC(0.027 higher),Acc(0.02 higher),Recall(0.02 higher),MAE(0.003 lower),and OOD generalization(0.056 higher).The results show that deep fully connected neural networks are good at predicting debris-flow susceptibility,thereby improving the accuracy and adaptability of debris-flow evaluation and providing a new way to evaluate debris-flow susceptibility.

关 键 词：泥石流灾害 易发性评价 深度学习算法 OOD泛化性验证 深度全连接神经网络 

分 类 号：P694[天文地球—地质学]