Multi-Feature Fusion-based Scene Classification Framework for HSR Images

To realize high-accuracy classification of high spatial resolution (HSR) images, this letter proposes a new multi-feature fusion-based scene classification framework (MF2SCF) by fusing local, global, and color features of HSR images. Specifically, we first extract the local features with the help of image slicing and densely connected convolutional networks (DenseNet), where the outputs of dense blocks in the fine-tuned DenseNet-121 model are jointly averaged and concatenated to describe local features. Second, from the perspective of complex networks (CN), we model a HSR image as an undirected graph based on pixel distance, intensity, and gradient, and obtain a gray-scale image (GSI), a gradient of image (GoI), and three CN-based feature images to delineate global features. To make the global feature descriptor resist to the impact of rotation and illumination, we apply uniform local binary patterns (LBP) on GSI, GoI, and feature images, respectively, and generate the final global feature representation by concatenating spatial histograms. Third, the color features are determined based on the normalized HSV histogram, where HSV stands for hue, saturation, and value, respectively. Finally, three feature vectors are jointly concatenated for scene classification. Experiment results show that MF2SCF significantly improves the classification accuracy compared with state-of-the-art LBP-based methods and deep learning-based methods.