摘要
滑坡现象对人类居住环境和城市建设的安全性构成了相当大的威胁,准确无误地识别滑坡就成为了防灾减灾领域内诸多研究课题的核心前置步骤.针对传统的滑坡体识别主要靠人工野外实地调查,危险性高、效率低、检测精度低的问题.首先,使用无人机倾斜摄影技术进行图像采集,其次,利用ContextCapture进行三维实景建模,最后通过深度学习技术可以对滑坡灾害区域进行图像分割,对滑坡发生的区域进行图像提取,从而达到识别的目的.选取山西省大同市云州区西坪村为研究区,通过基于深度学习的目标检测算法可以为滑坡体的精确识别定位提供数据支撑.通过对比YOLOV8-RCS-OSA、YOLOV3-tiny、YOLOV5n三种训练模型,最终选择YOLOV8-RCS-OSA进行目标检测,实验结果证明YOLOV8-RCS-OSA模型的mAP@0.5达到99.5%,mAP@0.5:0.95达到65.2%,对精确快速地获取滑坡信息对预测滑坡发生的可能性、减轻山体滑坡灾害影响、强化预防和治理效果都极具重要意义.
Abstract
Landslides pose a considerable threat to the safety of human settlements and urban construction,and therefore the ac-curate identification of landslides has become a central pre-requisite for many research projects in the field of disaster prevention and mitigation.Traditional landslide source identification mainly relies on manual field investigation,which is difficult,dangerous,inef-ficient and has low detection accuracy.In this paper,image acquisition is carried out by UAV tilt photography and then 3D model-ling is carried out by Context Capture in order to obtain the distance from the back wall of the landslide to reach the houses and cars of the inhabitants and to further assess the landslide hazard.Finally,through the application of deep learning-based landslide target detection algorithm to effectively detect the source of landslide objects,deep learning technology can quickly identify landslide haz-ards and image segmentation,extract the landslide occurred in the region of image recognition to further determine the landslide.In this paper,we compare YOLOV8-RCS-OSA,YOLOV3-tiny,YOLOV5n,and finally choose YOLOV8-RCS-OSA for the target de-tection of landslides and surrounding houses and cars.Datong City,Shanxi Province,Yunzhou District,Xiping Village as a research area,through the application of deep learning-based landslide target detection algorithms can effectively detect landslides,houses,cars are more suitable for landslide disaster prevention and control scenarios,to provide theoretical references and data support for the target's accurate identification and positioning.In this paper,by comparing the three training models of YOLOV8-RCS-OSA YOLOV3-tiny,YOLOV5n,we finally choose YOLOV8-RCS-OSA for landslide object source target detection,and through experi-ments,we proved that the mAP@0.5 of YOLOV8-RCS-OSA model for mudslide object source reaches 99.5%,mAP@0.5∶0.95 reaches 65.2%,and the overall detection effect is much better than YOLOv5n as well as YOLOv3-tiny algorithm models.Accurate and rapid determination of landslide information is very important for predicting the likelihood of landslides,reducing the impact of landslide disasters,improving prevention and control,optimising urban planning and protecting the ecological environment.
维普
万方数据