Deep Differential Coding for High-Resolution Remote Sensing Scene Classification

Abstract

Scene classification is one of critical tasks in the interpretation of high-resolution remote sensing (HRRS) imagery. Most of the existing methods focus on learning efficient feature representations. Recently, deep convolutional neural network (CNN) shows its effectiveness in extracting feature of images, which is widely used in remote sensing. In this paper, we propose a novel feature representation method which called deep differential coding (DDC) for remote sensing scene classification. Firstly, we get convolutional feature maps from image using the pre-trained CNN model. Then, the dense features are encoded through our proposed differential coding (DC). Finally, features are obtained by max-pooling and sum-pooling. We extract features from the whole remote sensing image and object region of every image by our proposed DDC and the final global feature is the combination of two kinds of features. We have performed extensive experiments on two public datasets and the results demonstrate that our proposed method has an excellent performance compared with state-of-the-art methods.