CSA-Net: An Adaptive Binary Neural Network and Application on Remote Sensing Image Classification

When deep neural networks are used to process remote sensing data, complex network structures, and parameters limit their real-time processing deployment in satellites. Binary Neural Networks (BNNs) can overcome this obstacle by reducing network complexity. However, the binarization process naturally leads to information loss, reducing the performance of the original neural networks. In this article, three modules are introduced to improve the performance of BNNs. First, we propose an improved channel shuffle (CS) module to enhance communication between different channels in feature maps, in which the grouped pointwise convolution is used to reduce the original calculation scale. Second, the squeeze and excitation (SE) module called the attention mechanism is inserted before each layer of the binary convolution operation to recognize the importance of different feature channels. Third, the adaptive linear scaling factor (ALSF) module is proposed to make the weights of the convolution kernel more suitable for the structure of BNNs, which further improves the performance of our network. With these three modules, our CSA-Net reaches or even outperforms the other state-of-the-art (SOTA) methods with a lower computational cost. It can achieve 65.35% Top-1 accuracy on the ImageNet, 90.11% accuracy on the CIFAR10, 89.58% accuracy on the WHU-RS19, and 90.86% accuracy on the UC-Merced, respectively. Meanwhile, compared with other mainstream remote sensing classification models, the number of our parameters is reduced by two or even three orders of magnitude, and the results confirm the effectiveness of the proposed method.