An Evaluation of Deep Domain Adaptive Networks for Remote Sensing Scene Image Classification

Abstract

The way to classify high-performance deep neural network depends on large-scale and high-quality labeled samples. However, it is difficult to obtain a large number of high-quality data sets in the field of remote sensing images. The most common solution is to use fine-tuned pre-trained neural network to study, which brings data deviation instead. At the same time, the remote sensing images which are in different time phases will be changed. It leads to violate the assumption between the training and test data. In this paper, we apply four deep adaptive networks to remote sensing image migration and classification experiments, and carried out an experimental evaluation of deep domain adaption for remote sensing image classification. Firstly, we selected two sets of public remote sensing optical scene classification data, two groups of data are compared in the above four networks, and the transfer accuracy under different network structures is obtained, and then we compared and analyzed the accuracy of different categories of data transfer classification. The experimental results show that the transfer accuracy of the four network structures using the transfer algorithm is much higher than the accuracy of the network with only transfer pre-trained model. The experimental results show that the accuracy of those four adaptive networks is higher, but the accuracy of different categories of data transfer classification is different, which can prove that the deep domain adaptive algorithm is also applicable to the remote sensing image transfer and can also confirm that the feasibility of mutual transfer between optical remote sensing data.