基于多特征卷积神经网络的哨兵二号影像地物分类  

Land Cover Classification of Sentinel-2 Image Based on Multi-feature Convolution Neural Network

作　　者：黄显培 孟庆祥[1] HUANG Xianpei;MENG Qingxiang(School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,Hubei,China)

机构地区：[1]武汉大学遥感信息工程学院,湖北武汉430079

出　　处：《应用科学学报》2023年第5期766-776,共11页Journal of Applied Sciences

基　　金：交通运输部2022年度交通运输行业重点科技项目(No.2022-3-2)资助。

摘　　要：空间分辨率为10 m的哨兵二号影像在原始的GoogLeNet中以影像的光谱值作为输入,没有将影像中的地物视为一个整体对象,为了利用影像的面向对象特征,提出了基于多特征的Object-oriented GoogLeNet网络结构。Object-oriented GoogLeNet在原有模型的基础上,引入了面向对象的光谱特征和形状特征,充分利用了不同地物间差异的形状特征进行分类。在武汉市及其周边的无云影像制作的数据集上,Object-oriented GoogLeNet模型的分类结果总体精度在GoogLeNet基础上提升了1.773%。结果表明,引入面向对象的特征模型在哨兵二号遥感影像分类中效果更好。The 10 m resolution Sentinel-2 images takes the spectral values of the image as input in the original GoogLeNet without treating the ground objects in the image as a whole.To leverage object-oriented features in Sentinel-2 remote sensing image classi-fication,this paper proposes an Object-oriented GoogLeNet network structure based on multiple features.Object-oriented GoogLeNet incorporates object-oriented spectral and shape features,and fully utilizes the shape features of differences between different ground objects for classification.On the data set of cloudless images in Wuhan and its surround-ing areas,the overall accuracy of the classification results of Object-oriented GoogLeNet model has increased by 1.773% compared to GoogLeNet.The results show that the model with object-oriented features enhances the classification performance of Sentinel-2 remote sensing images.

关 键 词：哨兵二号 GoogLeNet 深度学习 形状特征 土地利用分类 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]