基于CNN与SETR的特征融合滑坡体检测  

A landslide detection method using CNN-and SETR-based feature fusion

作　　者：李世琦 姚国清[1] LI Shiqi;YAO Guoqing(School of Information Engineering,China University of Geosciences(Beijing),Beijing 100083,China)

机构地区：[1]中国地质大学(北京)信息工程学院,北京100083

出　　处：《自然资源遥感》2024年第4期158-164,共7页Remote Sensing for Natural Resources

摘　　要：准确、及时地检测出滑坡体对减少山体滑坡自然灾害对人类生命和财产造成的威胁与损失具有重要意义。论文提出了一种基于卷积神经网络(convolutional neural network,CNN)与Set Transformer(SETR)的特征融合滑坡体检测方法。基于CNN的网络模型选择了全卷积网络(fully convolutional network,FCN)、U-Net和Deeplabv3+,基于Transformer的模型选择了SETR。首先对CNN网络模型在滑坡检测中的效果进行评价,然后在CNN网络模型的编码器部分引入SETR,并将SETR的输出融合到CNN的解码器结构中作为模型的整体输出。基于LandSlide4Sense数据集的实验结果表明,典型CNN融合SETR后有效改善了模型的检测效果,FCN,U-Net,Deeplabv3+模型在融合SETR后F1分数分别从0.6726,0.7273,0.6873提高到0.6869,0.7430,0.7055。因为滑坡与地形密切相关,以效果最好的U-Net模型为基准,在模型输入中引入数字高程模型之后F1分数从0.7325提高到0.7503。The accurate and timely detection of landslides is of great significance for reducing the threats to human life and properties,along with relevant losses,caused by landslides.This study proposed a landslide detection method using feature fusion based on convolutional neural networks(CNNs)and Segmentation Transformer(SETR).The CNN-based models utilized a fully convolutional network(FCN),U-Net,and Deeplabv3+,while the Transformer-based models used SETR.First,the landslide detection effects of the CNN-based models were evaluated.Then,SETR was introduced into the encoders of the CNN-based models,and the output of SETR was fused into the CNN decoder structure as the final output of the models.The experiments using the LandSlide4Sense dataset indicate that the fusion of typical CNNs with SETR can effectively improve the landslide detection effects.After SETR fusion,the FCN,U-Net,and Deeplabv3+models exhibited higher F1-scores,which increased from 0.6726,0.7273,and 0.6873 to 0.6869,0.7430,and 0.7055,respectively.Given the close relationship between landslides and terrain,a digital elevation model(DEM)was incorporated into the U-Net model,which outperformed other models.As a result,the F1-score of the model increased from 0.7325 to 0.7503.

关 键 词：CNN SETR 滑坡体检测 语义分割 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]