摘要
提出一种基于谷歌开源数据与支持向量机(SVM)结合的方法,以川西地区为试验区,实现遥感影像上滑坡的精准识别,识别对象多以与环境差异较大的新滑坡为主.基于开源谷歌影像建立川西历史滑坡数据库,包括历史滑坡1960处,将滑坡分为色调显著和纹理形态显著两类;利用方向梯度直方图进行特征提取训练SVM,利用训练好的模型进行滑坡二分类以及多分类识别.进行滑坡二分类时将样本数据集总数分为2000和4000以探讨样本数对SVM精度的影响,结果表明样本总数为2000时各模型识别精度上升,但滑坡的识别精度下降;进行滑坡多分类时将滑坡数据集分为色调显著型滑坡和纹理特征显著型滑坡,探讨模型在识别滑坡时纹理和色调特征的影响,再选用线性核(LN)SVM、多项式核(PL)SVM、高斯核(RBF)SVM、Sigmoid核(SIG)SVM 4种模型进行识别,模型的精度依次分别为0.69、0.72、0.77、0.70,LN-SVM和PL-SVM对识别滑坡纹理形态较为敏感,RBF-SVM和SIG-SVM更侧重于根据滑坡的色调特征识别滑坡.SVM可以应用于以谷歌地球影像为数据源的滑坡识别,且在滑坡识别时对滑坡的纹理特征更为敏感.
A method based on Google open source data and support vector machine(SVM)was proposed to achieve accurate landslide identification from remote sensing images in the west of Sichuan Province,and the recognition objects were mainly new landslides with great differences from the environment.A database of historical landslides in western Sichuan was established based on open source Google images,including 1960 historical landslides,and the landslides were classified into two categories:of a significant tone and a significant texture.Histogram of oriented gradient was used for feature extraction training SVM model,and the trained model was used for landslide binary classification recognition and landslide multi-classification recognition.The total number of sample data sets was divided into 2000 and4000 to explore the accuracy of SVM.The results showed that the identification accuracy of each model increased when the total number of samples was 2000,but the identification accuracy of landslide decreased.The landslide data set was divided into tone-significant landslide and texture-characteristic significant landslide in a multi-classification to explore the influence of texture and tone-characteristic on landslide recognition.The linear kernel(LN)SVM,polynomial kernel(PL)SVM,Gaussian kernel(RBF)SVM and Sigmoid kernel(SIG)SVM were selected for recognition.The accuracy of the four models were 0.69(LN-SVM),0.72(PL-SVM),0.77(RBF-SVM)and 0.70(SIG-SVM),respectively.LNSVM and PL-SVM of the four models were more sensitive to landslide texture morphology recognition,while RBF-SVM and SIG-SVM focused more on landslide identification according to their tonal characteristics.SVM was more sensitive to the texture characteristics of landslide recognition.
作者
宋雨洋
郝利娜
严丽华
王一
常浩
许强
SONG Yu-yang;HAO Li-na;YAN Li-hua;WANG Yi;CHANG Hao;XU Qiang(College of Earth Sciences,Chengdu University of Technology,Chengdu 610059,China;State Key Laboratory of Geohazard Prevention and Geo-environment Protection,Chengdu University of Technology,Chengdu 610059,China)
出处
《兰州大学学报（自然科学版）》
CAS
CSCD
北大核心
2022年第6期727-734,共8页
Journal of Lanzhou University(Natural Sciences)
基金
国家重点研发计划项目(2018YFC1505101,2021YFC3000401)
四川省地质灾害监测项目(510201202110324)。
关键词
滑坡
支持向量机
自动识别
谷歌影像
landslide
support vector machine
automatic identification
Google earth image
