All Adder Neural Networks for On-Board Remote Sensing Scene Classification

Directly performing remote sensing scene classification (RSSC) on satellites can effectively relieve pressures on data downlinks. However, existing convolutional neural network (CNN)-based methods with dense multiplication operations require a vast resource overhead and can hardly be deployed on resource-limited space platforms. In this article, an all adder neural network ( $\text{A}^{2}$ NN) with a generative-based hybrid knowledge distillation (GHKD) training strategy is proposed to solve this problem. In contrast to traditional CNNs, the $\text{A}^{2}$ NN discards all multiplication operations in the convolutional layers of the network and is instead completely composed of adder kernels; thus, this network has lower resource overheads than existing CNNs in hardware deployment. Compared with CNNs, the $\text{A}^{2}$ NN has a considerably lower RSSC accuracy; fortunately, this can be mitigated with the proposed GHKD training strategy. This strategy combines a hybrid KD method and a knowledge-matching-based generative learning method. The hybrid KD method utilizes the weight distribution knowledge in the adder neural network (ANN) to guide the high-performance classification knowledge in the CNN to be smoothly transferred to the $\text{A}^{2}$ NN, which consists entirely of adder kernels and has a considerably different structure than the CNN. The knowledge-matching-based generative learning method generates effective samples by matching the knowledge of the feature distribution of the real samples in the knowledge sources to augment the training database for hybrid KD. Extensive experiments and analyses on six public RSSC datasets show that the proposed GHKD training strategy significantly improves the RSSC performance of the proposed $\text{A}^{2}$ NN, achieving comparable performance to CNNs on most datasets.