基于残差网络的光学遥感图像场景分类算法

提出一种基于卷积神经网络中残差网络的遥感图像场景分类方法。本文方法在原网络模型中嵌入了跳跃连接和协方差池化两个模块,用于连接多分辨率特征映射和融合不同层次的多分辨率特征信息,并在3个公开的经典遥感数据集上进行了实验。结果证明,本文方法不仅可以将残差网络中不同层次的多分辨率特征信息融合在一起,还可以利用高阶信息来实现更具代表性的特征学习。与已有的分类方法相比,本文方法在场景分类问题上拥有更高的分类精度。

Scene Classification of Optical Remote Sensing Images Based on Residual Networks

This paper proposes a method for the scene classification of optical remote sensing images based on the residual network of convolutional neural networks.In the proposed method,two modules,i.e.,jump connection and covariance pooling,are embedded in the original network model to achieve multiresolution feature mapping and combine different levels of multiresolution feature information.Experiments are conducted on three open classical remote sensing datasets.Results show that the proposed method can fuse the multiresolution feature information of different levels in the residual network and use higher-order information to achieve more representative feature learning.The proposed method exhibits higher classification accuracy in the scene classification problem compared with the existing classification methods.