Scene classification of remote sensing image based on deep network grading transferring

Aiming at low precision of remote sensing image scene classification, a classification method DCNN_GT based on deep convolutional neural network (DCNN) and grading transfer (GT) is proposed. First, the DCNN model pre-trained on a large dataset (such as ImageNet) was transferred to a relatively smaller target dataset to fine-tune and the first-level classification features of images are extracted. Then, the similar images in the target dataset are clustered into several high-similarity categories, and the DCNN model of the first-level transfer is fine-tuned again on them and the second- level classification features are extracted. Then, the two-level classification features are encoded and fused, and the multi-kernel support vector machine (MKSVM) is used for scene classification. The experimental results in the common remote sensing datasets show that the average classification accuracy of the proposed method is improved by at least 1.83% compared with the one-time transfer and fine-tuning, especially the classification of confusable images is increased by at least 4%. In this paper, the DCNN is transferred gradually and is advanced to enhance the representation ability of the extracted image features, which makes the fusion features more recognizable. At the same time, the MKSVM is used to improve the generalization ability of the fusion features, so the classification result is better.