摘要
快速准确地自动识别滑坡对地质灾害普查工作、地质灾害管理以及风险评估具有重要意义。文章以谷歌地球(Google Earth)影像为数据源构建历史滑坡样本集,采用掩膜区域卷积神经网络Mask R-CNN目标检测算法实现了大范围滑坡信息自动检测,并以四川省丹巴县和甘肃省永靖县为研究区进行案例分析;同时结合影响滑坡发生的因子,以丹巴县为例,使用多层感知器模型获得研究区滑坡发生概率分区图,通过对比基于Mask R-CNN检测的滑坡灾害点在滑坡发生概率分区图中的分布情况验证滑坡检测的准确性。结果表明:山体滑坡检测的精确率为96%、召回率为85%、F 1值为0.90,且有79%的点分布在滑坡发生概率>80%的区域;黄土滑坡检测的精确率为98%、召回率为65%、F 1值为0.78;此方法可有效地提高地质灾害信息获取的自动化程度以及滑坡检测的精度。
Rapid and accurate automatic landslides identification is of great significance to geological hazard survey,management and risk assessment.In this paper,Google Earth images are utilized to construct a sample set of historical landslides,and the Mask R-CNN object detection algorithm is employed to detect massive landslides automatically.The proposed method is demonstrated by two case studies in Danba County,Sichuan Province and Yongjing County,Gansu Province.Considering the factors which affect the occurrence of landslides,taking Danba County as an example,this paper uses the multi-layer perceptron mode to obtain the occurrence probability map of landslides.The accuracy of landslides detection is further validated by comparing the distribution of landslides detected by Mask R-CNN against the landslides occurrence probability map.The results show that the precision of landslide detection is 96%,recall is 85%,F1 score is 0.90,and 79%of the points are distributed in the area where the probability of landslides occurrence is higher than 80%;the precision of loess landslides detection is 98%,recall is 65%,and F1 score is 0.78.This method can effectively improve the automation of geological hazard information acquisition and the accuracy of landslides detection.
作者
徐玲
刘晓慧
张金雨
刘震
XU Ling;LIU Xiaohui;ZHANG Jinyu;LIU Zhen(School of Surveying and Geo-Informatics,Shandong Jianzhu University,Jinan 250101,China)
出处
《山东建筑大学学报》
2023年第6期94-103,共10页
Journal of Shandong Jianzhu University
基金
山东省高校科技计划项目(J16LH05)
山东省自然科学基金项目(ZR2016DQ06)。
作者简介
徐玲(1998-),女,在读硕士,主要从事遥感图像处理与深度学习等方面的研究。E-mail:2020160102@stu.sdjzu.edu.cn;通讯作者:刘晓慧(1984-),女,副教授,博士,主要从事时空数据挖掘及空间信息技术在应急决策中的应用等方面的研究。E-mail:xhliu0512@163.com。
