PRUNING CONVOLUTIONAL NEURAL NETWORKS FOR REMOTE SENSING IMAGES

Deep convolutional neural networks (DCNNs) have shown excellent performance in remote sensing image processing. However, DCNNs contain many parameters and require many computational resources. It is very difficult to deploy DCNNs on spaceborne, airborne, or other mobile platforms with limited resources. It is necessary to compress the parameters of DCNNs. In this paper, we introduce a novel network pruning strategy that can prune redundant filters or weights of networks and preserve the performance of DCNNs in remote sensing image classification tasks. Unlike traditional pruning methods using a single criterion, this paper improves the pruning process based on the combined pruning method of multiple criteria. VGGNet and ResNet are used to validate our method. The dataset used in our experiments is the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset. The results demonstrate that our method is effective and can maintain the accuracy of a pruned model in remote sensing image classification tasks.