基于深度学习的地震诱发滑坡自动提取研究  

Automatic extraction of earthquake-induced landslides based on deep learning

作　　者：张明鹏 张帅 吕运鸿 ZHANG Ming-peng;ZHANG Shuai;LV Yun-hong(College of Civil Engineering and Architecture,Zhejiang University,Hangzhou 310058,Zhejiang,China)

机构地区：[1]浙江大学建筑工程学院,浙江杭州310058

出　　处：《地基处理》2024年第3期242-249,共8页Journal of Ground Improvement

摘　　要：地震诱发滑坡的提取是震区工程风险评价的基础,对震区的灾后重建工作具有重要意义。传统的机器学习方法需要复杂的数据预处理及设计特征工程工作,而深度学习方法则可以直接输入图像数据进行端对端学习,从而实现对地震诱发滑坡的自动提取。本文探究了U-Net与LinkNet两种卷积神经网络在地震诱发滑坡自动提取任务中的应用。研究表明U-Net和LinkNet在地震诱发滑坡自动提取应用中具有巨大的潜力,召回系数、F1分数、精确度均达到0.8以上。相较之下,LinkNet网络结构在滑坡的自动提取任务中整体性能优于U-Net,召回系数、F1分数、精确度分别提升2.0%~8.0%、4.0%~8.0%、8.0%~10.0%。U-Net网络结构适用于简单背景图像中大型滑坡的提取,而LinkNet网络结构则更适用于复杂背景图像中的滑坡及简单背景图像中小型滑坡的提取。Extracting earthquake-induced landslides is foundational for engineering risk assessment in seismic regions and holds significant importance for reconstruction efforts.Traditional machine learning methods require complex data preprocessing and feature extraction.In contrast,deep learning methods enable direct input of image data for end-to-end learning,facilitating automatic extraction of earthquake-induced landslides.This work investigates the automatic recognition of regional earthquake-induced landslides using U-Net and LinkNet convolutional neural networks.Experimental results show that U-Net and LinkNet exhibit significant extraction capabilities in automatically identifying earthquake-induced landslides.The recall coefficient,F1-score and precision of evaluation indexes all exceed 0.8.LinkNet outperforms U-Net overall,with improvements in the recall coefficient(2.0%~8.0%),F1-score(4.0%~8.0%),and precision(8.0%~10.0%).The U-Net network structure is well-suited for extracting large landslides in simple background images,whereas the LinkNet network structure is better suited for extracting landslides in complex background images and small landslides in simple background images.

关 键 词：地震 滑坡 卷积神经网络 深度学习 自动提取 遥感影像 

分 类 号：P237[天文地球—摄影测量与遥感]