High-Resolution Remote Sensing Scene Classification Using Improved LBP and SDSAE

Abstract

Classifying high-resolution remote sensing scene images with high accuracy is the challenging issues. The key of scene classification is to find effective features of the scene image. Low-level visual feature methods, such as local binary pattern (LBP), assume the same type of scene should share certain statistically holistic attributes and have demonstrated their efficiency on scene classification. In this paper, we propose an effective LBP variant, called IRELBP, which use radial difference and angular difference as its difference based descriptors to represent HRRS scene images. Due to the fact that low-level features cannot represent more meaningful semantic information, we extract features from a Stack Denoising Sparse Autoencoder (SDSAE) to obtain more meaningful hierarchical features. Both the global features and hierarchical features are encoded by Fisher Vector, and then they are concatenated into a discriminative representation, which is fed into the SVM classifier for training or testing. We perform comprehensive experiments on two remote sensing scene classification benchmarks: UC-Merced dataset and the recently introduced large scale aerial image dataset (AID). The result demonstrates that our proposed combination method can provide effective and discriminate feature representation and outperforms the state-of-the-art methods in HRRS scene image classification.