**5. Conclusions**

In this paper we have proposed a robust image classification algorithm based on deep learning integrated with binary coding and Sinkhorn distance. Taking into account the characteristics of hand-crafted features and deep features, we combine their advantages and supplement the deep features with the statistical texture features to fully describe the image. In order to remove redundant information from the fused features and train the model quickly and efficiently, we introduced the Sinkhorn loss where an entropy regularization term plays a key role. In this paper, experiments are carried out on two classic texture datasets and five remote sensing classification datasets. The experimental results show that compared with the ResNet-50, the proposed two stream model DBSNet can improve the overall performance when achieving image classification tasks. In addition, compared with the classic classification algorithms for remote sensing scene classification, the algorithm DBSNet can still provide better results. In the future, we will study how to combine the traditional feature extraction framework with the deep learning framework so that they guide and improve each other.

**Author Contributions:** Conceptualization, C.H.; Funding acquisition, C.H.; Investigation, Q.Z.; Methodology, C.H. and Q.Z.; Writing—original draft, Q.Z.; Writing—review & editing, T.Q., D.W. and M.L.

**Funding:** This research was funded by the National Key Research and Development Program of China (No. 2016YFC0803000), the National Natural Science Foundation of China (No. 41371342, No. 61331016), and the Hubei Innovation Group (2018CFA006).

**Conflicts of Interest:** The authors declare no conflict of interest.
