基于地形特征融合的卷积神经网络滑坡识别

Convolutional Neural Network Landslide Recognition Based on Terrain Feature Fusion

滑坡严重威胁着人民群众的生命财产安全。完整、准确的滑坡编录图是研究滑坡的重要资料。深度卷积神经网络方法由于众多优势而备受关注,然而卷积神经网络结构复杂,需要大量的训练样本,制约了其在滑坡制图上的发展。提出了融合地形特征的卷积神经网络建模方法。首先在遥感影像上叠加地形因子构建新的滑坡样本,然后设计提取并融合空间与光谱特征的轻量级卷积神经网络(FF-CNN),最后训练最优模型进行滑坡识别。在四川汶川地区进行的消融实验证明:在空间特征基础上融合光谱特征的FF-CNN模型滑坡识别评价指标F1分数和平均交并比(MIoU)分别提高0.0202和0.0144;在遥感影像上叠加地形因子后,FF-CNN模型滑坡识别评价指标F1分数和MIoU值分别提高0.0664和0.0482。在湖北省三峡库区和四川省都江堰市虹口乡的实验说明FF-CNN模型表现出较强的适用性和迁移能力,在滑坡识别上具有较大潜力。

Landslides seriously threaten the safety of people s lives and property.A complete and accurate landslide inventory map is important for the study on landslides.The deep convolutional neural network method has attracted great attention due to its numerous advantages.However,the convolutional neural network has a complex structure and requires lots of training samples,which restrict the development of this technology in landslide recognition.A convolutional neural network modeling method incorporating terrain features was proposed.Firstly,a new landslide sample is constructed by superimposing terrain factors on remote sensing images,then a lightweight convolutional neural network that extracts and fuses spatial and spectral features(FF-CNN)is designed,and finally the optimal model is trained for landslide recognition.The ablation experiments in Wenchuan area of Sichuan show that after fusing spectral features on the basis of spatial features,the F1 score and MIoU of FF-CNN model are increased by 0.0202 and 0.0144,respectively;after superimposing the terrain factor on the remote sensing image,the F1 score and MIoU of FF-CNN model are increased by 0.0664 and 0.0482,respectively.The experiments in Three Gorges reservoir area of Hubei province and Hongkou town of Dujiangyan city,Sichuan province,show that FF-CNN model has strong applicability,migration ability and great potential in landslide recognition.