Transfer Learning for Classification of Optical Satellite Image

Abstract

Deep convolutional neural network (DCNN) has achieved great success in the classification of natural images, but it requires numerous labelled data for training. In the absence of a large number of optical satellite images and labelled data, how to guarantee the effect of classification of the optical satellite images with DCNN? In this case, this paper has discussed how to fine-tune a pre-trained DCNN in a layer-wise manner by transfer learning. In our experiment, DCNN is pre-trained with ImageNet which is a large labelled dataset of natural images, and then optical remote sensing images are used to fine-tune the learnable parameters of pre-trained DCNN. The experimental results show that transfer learning is feasible to deal with the above problem. In the process of transfer training, if the second half of the layers are fine-tuned, compared with the fine-tuning of the entire network, the almost same accuracy can be achieved, but the convergence is more rapid. The experimental results provide a solution for how to achieve the incremental classification performance in practical applications.