基于DeepLabv3+算法的黄土地区地质灾害隐患形变区识别技术研究
Research on deformation identification technology of suspected geological disasters in loess area based on DeepLabv3+
魏统彪 1石鹏卿 1高子雁 1周小龙1
作者信息
- 1. 甘肃省地下水工程及地热资源重点实验室,甘肃兰州 730050;甘肃省地质环境监测院,甘肃兰州 730050
- 折叠
摘要
深度学习在信息提取方面具有非常大的优势,因此选用深度学习对兰州市地质灾害隐患形变区进行提取.由于DeepLabv3+算法中Xception结构复杂,计算机运行时占用内存较多,因此,对比DeepLabv3+算法3种不同主干网络ResNet101、MobileNet以及DRN,选择最优的主干网络作为DeepLabv3+算法主干网络,并且将选取的最优主干网络DeepLabv3+算法提取的地质灾害隐患形变区和较为流行的算法PSPnet、FCN32s提取的地质灾害隐患形变区进行对比.结果显示:改进的算法一定程度上可以提高地质灾害隐患形变区的识别精度,效果优于FCN32s、PSPnet算法所提取的地质灾害隐患形变区的识别精度,其中IOU比FCN32s高12.98％,比PSPnet高1.40％,preci-sion 比 FCN32s 高 10.79％,比 PSPnet 高 0.15％,recall 比 FCN32s 高 6.10％,比 PSPnet 高 1.50％.研究表明改进的算法在地质灾害隐患形变区识别方面更有优势.
Abstract
Deep learning has great advantages in information extraction,so this paper chooses deep learning to extract deformation areas of geological hazards in Lanzhou City.Due to the complex Xception structure and high resource consumption in DeepLabv3+algorithm,this paper compares three different backbone networks of DeepLabv3+algorithm,ResNet101,MobileNet and DRN,and selects the optimal backbone network as the backbone network of DeepLabv3+algorithm.Moreover,the deformation area of geological disaster hidden danger extracted by DeepLabv3+algorithm is compared with the popular algorithms PSP-net and FCN32s,the deformation areas extracted from geological hazards are compared.The results show that to a certain extent,the improved algorithm can improve the identification accuracy of hidden geological disaster deformation area,and the effect is better than that of FCN32s and PSPn et al gorithms.The IOU is 12.98％higher than FCN32s and 1.40％higher than PSPnet.The precision is 10.79％higher than FCN32s,0.15％higher than PSPnet,and the recall is 6.10％higher than FCN32s and 1.50％higher than PSPnet.The results show that the proposed algorithm has more advantages in the identification of hidden deforma-tion areas of geological disasters.
关键词
深度学习/地质灾害/DeepLabv3+/隐患识别Key words
Deep learning/Geological disasters/DeepLabv3+/Hazard identification引用本文复制引用
基金项目
甘肃省自然资源厅科技创新项目(202257)
出版年
2024
NETL
NSTL
万方数据