基于双通道残差网络的泥石流沟谷孕灾风险预测  

Debris Flow Ravine Hazard Prediction Based on Two-Way Residual Network

作　　者：袁若浩 王保云 YUAN Ruohao;WANG Baoyun(School of Mathematics,Yunnan Normal University,Kunming 650500,China;Key Laboratory of Modeling and Application of Complex Systems in Universities of Yunnan Province,Kunming 650500,China)

机构地区：[1]云南师范大学数学学院,云南昆明650500 [2]云南省高校复杂系统建模及应用重点实验室,云南昆明650500

出　　处：《贵州大学学报（自然科学版）》2023年第6期78-85,115,共9页Journal of Guizhou University:Natural Sciences

基　　金：国家自然科学基金资助项目(61966040)。

摘　　要：云南作为泥石流受灾最严重的省份之一,每年均会遭受重大损失。为了应对这种突发性灾害,本文基于DCHNNet(dual-channel hybrid neural network)提出了一个基于双通道的改进残差结构的卷积神经网络——双通道残差网络(two-way residual network,TWRNet)。该网络能够广泛应用于泥石流沟谷图像的潜在危险性排查,实现泥石流灾害的预警。TWRNet首先采用切片的方式对数字高程(digital elevation model,DEM)数据和遥感数据分开处理,并使用改进的残差结构进行特征提取;然后将特征进行融合,并使用通道注意力机制SE(squeeze-and-excitation networks)模块进行通道增强;最后给出泥石流沟谷的分类结果。在训练过程中,本文使用了交叉熵和焦点损失构成的联合损失函数。实验结果表明,TWRNet在泥石流沟谷识别方面达到了最高89.28%的识别率和87.50%的召回率,模型性能良好。使用图像学习沟谷特征的方法来进行泥石流孕灾沟谷的识别是可行的。Yunnan,as one of the provinces most affected by debris flow disasters,suffers significant losses every year.In response to this sudden natural hazard,the author proposes a two-way residual network(TWRNet),based on the dual-channel hybrid neural network(DCHNNet).TWRNet is capable of extensively assessing the potential hazards of debris flow ravines through image analysis,facilitating early warning systems for debris flow disasters.TWRNet first separates and processes digital elevation model(DEM)data and remote sensing data using a slice-based approach,extracting features with an improved residual structure.Subsequently,the features are fused and enhanced using the channel attention mechanism called the SE module.Finally,the network provides classification results for debris flow ravines.The training process utilizes a joint loss function consisting of cross-entropy and focal loss.Experimental results demonstrate that TWRNet achieves a maximum recognition rate of 89.28%and a recall rate of 87.50%for debris flow ravine identification,exhibiting excellent model performance.Employing the method of image-based learning for the recognition of ravine features associated with debris flow hazards is feasible.

关 键 词：泥石流 分类 卷积神经网络 预测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] P642.23[自动化与计算机技术—控制科学与工程]