Multilayer feature fusion using covariance for remote sensing scene classification

Abstract

Remote sensing images are obtained by electromagnetic measurement from the terrain of interest. In high-resolution remote sensing imageries extraction measurement technology plays a vital role. The scene classification is one of the interesting and challenging problems due to the similarity of image structure and the available HRRS image datasets are all small. Training new Convolutional Neural Networks (CNN) using small datasets is prone to overfitting and poor attainability. To overcome this situation using the features produced by pre-trained convolutional nets and using those features to train an image classifier. To retrieve informative features from these images we use the existing Alex Net, VGG16, and VGG19 frameworks as a feature extractor. To increase classification performance further makes an innovative contribution fusion of multilayer features obtained by using covariance. First, to extract multilayer features, a pre-trained CNN model is used. The features are then stacked, downsampling is used to stack features of different spatial dimensions together and the covariance for the stacked features is calculated. Finally, the resulting covariance matrices are employed as features in a support vector machine classification. The results of the experiments, which were conducted on two difficult data sets, UC Merced and SIRI-WHU. The proposed Staked Covariance method consistently outperforms and achieves better classification performance. Achieves accuracy by an average of 6 % and 4 %, respectively, when compared to corresponding pre-trained CNN scene classification methods.