A Network Pruning Method for Remote Sensing Image Scene Classification

Deep convolutional neural networks have been widely used to improve remote sensing image scene classification performance. However, most of these networks include many parameters and need many computational resources. Which hinders the applications of these networks when facing the satellite, plane or other mobile platforms. In this paper, we introduce a network pruning framework which can reduce the size of the network model and maintain the classification accuracy. In this framework, we train the pruned model using both the original unpruned model's output and training dataset. Which can learn more information than retrain using dataset only. In our experiments, we evaluate our method for remote sensing scene classification on NWPU-RESISC45 dataset. The results demonstrate that our method was effective and maintained the model classification accuracy.