摘要： 为充分利用遥感图像的场景信息，提高场景分类的正确率，提出一种基于空间特征重标定网络的场景分类方法。采用多尺度全向高斯导数滤波器获取遥感图像的空间特征，通过引入可分离卷积与附加动量法构建特征重标定网络，利用全连接层形成的瓶颈结构学习特征通道间的相关性，对多尺度空间特征进行权重筛选以实现特征重标定，并结合卷积神经网络训练得到最终的分类结果。实验结果表明，该方法在UCM_LandUse与机载SAR图像数据上的分类正确率分别达到94.76%和95.38%，与MNCC、MS-DCNN、PCA-CNN等算法相比，其遥感图像分类精度与泛化能力显著提升。

关键词: 遥感图像, 场景分类, 多尺度空间特征, 特征重标定, 卷积神经网络

Abstract: In order to make full use of the scene information of remote sensing image and improve the accuracy rates of scene classification,this paper proposes a scene classification method based on spatial feature recalibration network.The Multi-scale Omnidirectional Gaussian Derivative Filter(MOGDF) is constructed to obtain multi-scale spatial features of remote sensing images.Then,a feature recalibration network is constructed by introducing separable convolution and additional momentum methods,and the bottleneck structure is formed by using the fully connected layer to learn the correlation between the feature channels.The multi-scale spatial features are weighted to achieve the recalibration of features.Finally,combined with the Convolutional Neural Network(CNN) training,the classification results are obtained.Experimental results on UCM_LandUse and airborne SAR image datasets show that the accuracy rates of the proposed method for remote sensing image classification reach 94.76% and 95.38%,respectively,and compared with algorithms such as MCNN,MS-DCNN,PCA-CNN and so on,the accuracy and generalization ability of remote sensing image classification are significantly improved.

Key words: remote sensing image, scene classification, multi-scale spatial features, feature recalibration, Convolutional Neural Network(CNN)

中图分类号: 

• TP751