Towards Exploiting Convolutional Features for Remote Sensing Images Scene Classification

Abstract

The developments in deep learning have shifting the research directions of remote sensing image scene understanding and classification from pixel-to-pixel handcrafted methods to scene-level image semantics analysis for scene classification tasks. The pixel-level methods rely on handcrafted methods in feature extraction, which yield low accuracy when these extracted features are fed to Support Vector machine for scene classification task. This paper proposes a generic extraction technique that is based on convolutional features in the context of remote sensing scene classification images. The experimental evaluation results with convolutional features on public datasets, Whu-RS, Ucmerced, and Resisc45 attain a scene classification accuracy of 92.4%, 88.78%, and 75.65% respectively. This demonstrate that convolutional features are powerful in feature extraction, therefore achieving superior classification results compared to the low-level and mid-level feature extraction methods on the same datasets.