Remote Sensing Image Classification Based on Lightweight Network and Pruning

Abstract

In recent years, deep neural network has continuously improved the performance of remote sensing image classification. Though The deep neural network model is powerful, it is difficult to deploy on resource-constrained hardware platforms such as mobile terminal devices and embedded devices due to the large number of network parameters. In this paper, a novel lightweight network (LW-Net) model and a network pruning method are used for remote sensing image classification. This LW-Net model adopts a net block unit to obtain more characteristic graphs with less computational complexity. The proposed network pruning method uses the sparsity regularization on the influence factor in BN layer to automatically identified and pruned unimportant channels to make the model structure simpler. Experimental results demonstrate compared with traditional deep neural networks, the proposed model with the network pruning method can greatly reduce the computational complexity and model parameters with a little accuracy loss.