Classification For Hyperspectral Remote Sensing Image Based On Deep Learning

The hyperspectral image not only has high spatial resolution, but also has high spectral resolution which usually is nanoscale. Comparing with the traditional multispectral image, the hyperspectral image has rich spatial and spectral information and more suitable for classification. In recent years, with the growing of the application for hyperspectral remote sensing, the technology is also increasing. As a foundation of its use in many fields, hyperspectral remote sensing image classification is an important way for people to extract useful information from the hyperspectral data. In this paper, we put the classification of the hyperspectral remote sensing image as the goal and mainly studied the following contents:Firstly, there are fewer labeled samples in hyperspectral images because of the difficulty to obtain and collect the labeled samples. In order to deal with this problem, a half supervised classification model based on the stacked denosing autoencoder was proposed: in the first, the vast unlabeled samples are used to unsupervised training the network, then the deep and abstract characteristics could be extracted by the network; then, combing the Softmax classifier with the stack denoising autoencoder together to form a deep classification network and use the labeled samples to supervised training the network. This paper also studied the effect of the deep and the number of the hidden layer for the classification.Then, most current classification algorithms ignore the using of the spatial information and only make use of the spectral information. To improve the usage of the spatial information, spatial-spectral classification methods based on edge-preserving filter and deep learning were proposed. The methods combing with the spatial information before classification(BS_EPF_SDAE), after classification(SDAE_EPF), and before-after classification(BS_EPF_SDAE_EPF) were discussed respectively. BS_EPF_SDAE combined band selection, edge-preserving filter and stacked denoising autoencoder together, which can not only reduce the amount of calculation of the hyperspectral data, but also can combine the spatial and spectral information together. And the deep learning network was used to obtain the useful high-level features and do the classification. The experimental results show this method can effectively remove the noisy points and can improve the classification accuracy. SDAE_EPF used an edge-preserving filter to post-process the initial classification results, which can make the adjacent samples have similar values in each probability maps and maintain the edge information which means the results can be well aligned with the real object. BS_EPF_SDAE_EPF combined BS_EPF_SDAE and SDAE-EPF together, which can makes full use of the spatial and spectral information to reduce the noisy points and improve the accuracy of the classification.