基于多层特征信息融合的滑坡图像分割模型  

Landslide image segmentation model based on multi-layer feature information fusion

作　　者：张银胜 陈戈[2] 段修贤 童俊毅 单梦姣 单慧琳 ZHANG Yinsheng;CHEN Ge;DUAN Xiuxian;TONG Junyi;SHAN Mengjiao;SHAN Huilin(Jiangsu Integrated Circuit Reliability Technology and Testing System Engineering Research Center,Wuxi University,Wuxi 214105,China;School of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;Key Laboratory of Intelligent Support Technology for Complex Environment in Nanjing University of Information Science and Technology,Ministry of Education,Nanjing 210044,China)

机构地区：[1]无锡学院江苏省集成电路可靠性技术及检测系统工程研究中心,江苏无锡214105 [2]南京信息工程大学电子与信息工程学院,江苏南京210044 [3]南京信息工程大学复杂环境智能保障技术教育部重点实验室,江苏南京210044

出　　处：《测绘学报》2024年第11期2201-2212,共12页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(62071240,62106111);无锡市科技发展资金“太湖之光”科技攻关(基础研究)项目(K20231004)。

摘　　要：滑坡对人类生存环境造成严重的危害,人工识别滑坡区域的方式比较耗时且隐蔽区域不易被探测,而利用遥感图像识别滑坡区域,能够准确快速地实现滑坡灾害预警和救援。随着深度学习的快速发展,语义分割已经广泛应用于滑坡遥感图像识别领域。针对当前滑坡图像分割模型容易出现错误识别、图像边缘信息丢失等问题,本文提出了一种多层特征信息融合的滑坡分割模型MLFIF-Net,该模型以MobileNetv3-Small为主干网络,提高模型对滑坡图像的特征提取能力,同时构建级联带状空间金字塔池化模块增强滑坡图像的纹理特征,获取多尺度信息,利用高效通道注意力模块关注图像特征,设计了多层特征信息融合结构增强图像的边缘信息,从而提升模型的分割效果。试验结果表明,本文模型在贵州毕节市滑坡数据集上的准确率为96.77%,类别平均准确率为95.61%,平均交并比达到了87.69%,与SegNet等6种分割模型相比,其分割精度较为优异,能够准确识别目标区域,突出滑坡图像边缘细节。Landslide cause serious harm to human living environment.The method of manually identifying the landslides is time-consuming and the hidden area is not easy to detect.The use of remote sensing image to identify the landslides can accurately and quickly realize the landslide disaster warning and rescue.With the rapid development of deep learning,semantic segmentation has been widely used in the field of landslide remote sensing image recognition.Aiming at the problems such as error recognition and image edge information loss in the current landslide image segmentation model,this paper proposes a landslide segmentation model MLFIF-Net,which integrates multi-layer feature information fusion.The model uses MobileNetv3-Small as the main trunk network to improve the feature extraction ability of the model.At the same time,a cascade spatial pyramid pool module is constructed to enhance the texture features of landslide images and obtain multi-scale information.An efficient channel attention module is used to focus on image features,and a multi-layer feature information fusion structure is designed to enhance the edge information of images,so as to improve the segmentation effect of the model.The experimental results show that the accuracy of the proposed model on the landslide data set of Bijie city,Guizhou province is 96.77%,the average accuracy of the class is 95.61%,and the average interaction ratio is 87.69%.Compared with SegNet and other six segmentation models,its segmentation accuracy is better,and it can accurately identify the target area and highlight the edge details of the landslide image.

关 键 词：语义分割 遥感图像 滑坡 金字塔池化 注意力模块 特征信息融合 

分 类 号：P237[天文地球—摄影测量与遥感]