Branch Feature Fusion Convolution Network for Remote Sensing Scene Classification

Convolutional neural networks (CNNs) have outstanding advantages in the classification of remote sensing scenes. Deep CNN models with better classification performance typically have high complexity, whereas shallow CNN models with low complexity rarely achieve good classification performance for remote sensing images with complex spatial structures. In this article, we proposed a new lightweight CNN classification method based on branch feature fusion (LCNN-BFF) for remote sensing scene classification. In contrast to a conventional single linear convolution structure, the proposed model had a bilinear feature extraction structure. The BFF method was utilized to fuse the feature information extracted from the two branches, which improved the classification accuracy. In addition, combining depthwise separable convolution and conventional convolution to extract image features greatly reduced the complexity of the model on the premise of ensuring the accuracy of classification. We tested the method on four standard datasets. The experimental results showed that, compared with recent classification methods, the number of weight parameters of the proposed method only accounted for less than 5% of the other methods; however, the classification accuracy was equivalent to or even superior to certain high-performance classification methods.