Deep feature representations for high-resolution remote sensing scene classification

Learning powerful feature representations is of great importance in order to improve the accuracy of high-resolution remote sensing imagery (HRRSI) scene classification. Generally, traditional methods mainly focus on low-level features. Though these feature representations can achieve satisfactory performance to some extent, the performance improvement is small. More importantly, these low-level feature representations are hand-crafted features, and it is difficult to design such a holistic feature representation. Recently, convolutional neural networks (CNN) has achieved remarkable performance on several natural benchmarks. In this paper, we investigate the extraction of the deep feature representations based on the pre-trained CNN architectures for scene classification tasks. More specially, we first evaluate the pre-trained CNN architectures on a public scene dataset and then fine-tune the CNN that performs best on the dataset using the target HRRSI dataset to learn dataset-specific features. Extensive experiments on two publicly available scene datasets indicate that the deep feature representations can achieve state-of-the-art performance.