Remote sensing image scene classification via multi-feature fusion

Abstract

In this papers, classification of remote sensing image scene is investigated. A scene classification approach based on multi-feature fusion has been proposed. In the proposed approach, three types of features are extracted. Specifically, extended multi-attribute profile (EMAP)-based texture feature, saliency-based shape feature and color ones. The texture features are extracted by EMAP. Furthermore, the Hu invariant moments are extracted from the saliency map, where the saliency map is obtained by frequency-tuned saliency detection. Meanwhile, the color moments are extracted as the color features from the image scenes. As for EMAP-based features, dimension reduction via principal component analysis (PCA) is first performed and combined with other two types of features to form a compact feature representation. Finally, support vector machine (SVM) is employed to classify the remote sensing image scenes. The experiments on the two challenging image scene datasets are performed to show that the proposed method is simple, yet efficient to implement, comparing with the state-of-the-arts.