Abstract：Aiming at the problem that the traditional scene classification methods not able to accurately express the rich semantic information of high-resolution remote sensing images， a high-resolution image scene classification method based on convolutional neural network is proposed. The method is roughly divided into three steps. The first step is performing convolution operations according to different convolution windows to extract low-level features such as color， texture and shape. The second step is using the pooling layer to filter these low-level features to obtain important features. The third step is recoganizing the extracted features to form high-level semantic features for scene classification. The experiment contains a comparison of the classification effects of three different convolution kernels. In addition， data augmentation， regularization and dropout are used to improve the adaptability of the model to new samples and solve the problem of over-fitting. The method performs well in the experiments conducted. It achieves an accuracy of 88.47% on the WHU-RS19 dataset， which significantly improves the classification accuracy compared with traditional scene classification methods.
2019, Vol. 53 
