Research on remote sensing image classification method using two-stream convolutional neural network

Abstract

Owing to the lack of remote sensing image data set and no regional pertinence in terms of characteristics for classification, we have published the remote sensing image of some areas in Haikou City, Hainan Province, and made a HN-7 dataset, which has the regional characteristics specific of Hainan Province. The HN-7 dataset consists of seven classes, of which the construction site and dirt road categories appear in the public remote sensing dataset for the first time. Owing to the limited quantity of the HN-7 dataset, we decided to train a small convolutional neural network from scratch for the classification task, by using a three-layer two-stream network for improving the accuracy of the neural network model. Our model achieved 98.57% accuracy on the test set. We compared the accuracy of four common networks trained on HN-7, and the results showed that our model achieves the best performance.