摘要
山区多发沟谷型泥石流,而由于山区地形崎岖,导致无法开展大面积的泥石流危险性评价工作。本文使用遥感数据、DEM(Digital Elevation Model)数据以及岩性、土壤、植被数据,构建了一个基于多源数据,能快速进行大面积排查工作的卷积神经网络模型RSDNet(Residual-Shuffle-Dense residual Net)。该模型首先使用最大池化改进的残差结构对各类不同数据进行浅层特征提取,然后使用通道重排以加强各类数据底层特征间的关联性,接着使用密集残差结构对底层特征作进一步的特征提取,学习各类特征间的相互作用对潜在泥石流危险性的影响,最后根据待评价沟谷与已发生过泥石流沟谷的相似度给出沟谷的潜在危险性。在训练过程中,使用了交叉熵和基于焦点损失改进的联合损失函数,使模型能更好地区分各类沟谷的形态特征和致灾特征。RSDNet在沟谷分类任务上可达到89.7%的精确率。在对怒江州全境沟谷进行潜在危险性评价的任务中,132条历史泥石流沟谷有122条被模型判断为高危险或极危险。结果表明模型性能良好,为沟谷泥石流的危险性评价提供了新思路。
Gully debris flows frequently occur in mountainous regions.Hazard investigation in large area is always hampered by the rugged mountains.In this paper,a convolutional neural network named ResidualShuffle-Dense residual Net(RSDNet)based on remote sensing,DEM,soil,lithology,and vegetation data was proposed to conduct large-scale territorial surveys.First,shallow features were extracted by modified residual structure using maximize pooling.Then,feature fusion was conducted to strengthen the correlation between the underlying features of various data.Next,dense residual structure was applied to make further feature extraction from underlying features and identify the impact of interaction between various features on potential debris flow hazard.Finally,the potential hazard level of a valley was given.During the training process,a joint loss function based on cross entropy loss and modified focal loss was used to make the model better distinguish the morphological and disaster-causing characteristics of various valleys.In this study,the valley classification achieved a precision rate of 0.92 by using RSDNet.In the potential hazard assessment of all valleys in Nujiang Prefecture,122 out of 132 historical debris flow valleys were judged to be dangerous or very dangerous.Results indicate that the proposed model performs well,and this work would offer new ideas for potential debris flows hazard assessment.
作者
徐繁树
王保云
韩俊
XU Fanshu;WANG Baoyun;HAN Jun(School of Information,Yunnan Normal University,Kunming 650500,China;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)
出处
《地球信息科学学报》
CSCD
北大核心
2023年第3期588-605,共18页
Journal of Geo-information Science
基金
国家自然科学基金项目(61966040)。
关键词
泥石流
灾害预测
山区沟谷
多源数据融合
怒江州
卷积神经网络
残差模块
焦点损失
debris flow
hazard prediction
mountainous valleys
multisource data fusion
Nujiang Prefecture
convolutional neural network
residual block
focal loss
