High Resolution Remote Sensing Image Classification Using Hybrid Ensemble Learning

The classification of the high resolution remote sensing image has become very complex as the resolution improves. Recently, deep learning has been used successfully for high resolution image classification. But it is hard to classify different objects with a large number of pixels, and a small number of pixels at the same time with a single learning model. In this paper the hybrid ensemble learning method is proposed, which combines three kinds of networks: fully connected network, convolutional neural network, and fully convolutional network, to obtain high classification performance. Firstly, the fully connected network is carried out using pixel brightness as features for classification. Secondly, the object-oriented segmentation was used, and the convolutional block was extracted from the center of gravity and the convolutional neural network was used for classification. Thirdly, images were clipped into image blocks, and fully convolutional network U-Net was used with one versus all. Finally, the hybrid ensemble learning method is used, in which a fully connected network is trained for probabilistic combination using the classification results of the three base classifiers. The proposed method was tested with a large number of experiments on Vaihingen aerial image dataset, and the results showed the hybrid ensemble learning method has best performance compared with the single classical neural network and the deep learning method.