Researches On Data Compression And Classification Of Hyperspectral Images

Hyperspectral images hold the characteristics of high dimensionalities and large numbers of pixels. Researches are mainly focused on feature selection, feature extraction, pattern classification, etc. The two most important aspects in image processing are feature learning and analysis of key data, because of the huge image data points and redundancy among them. On one hand, feature learning algorithms are widely used, such as PCA, LDA and increasingly popular deep-learning based methods. On the other hand, data compression is not famous, yet has some valuable algorithms, such as kNN-based data selection method and Nystrom-based method. This paper gives a combination of kNN-based, neural-network based method and deep-learning algorithm to select key data in hyperspectral images. Then the selected data are labelled as training set to guide the modelling process and improve the classification results. The last work in this paper shows the conversion of CNN, which is widely used in 2D image recognition, to a situation that classification can be viable in hyperspectral images. The main researches can be summarized as follows:1. A multi-layer-based data compression and classification method is proposed for the selection of key data in hyperspectral images. According to the demands of users, the ratio of compression in each layer can be self-adjusted, until the selected key data in the last layer can well represent the original images. Then the key data is labelled as training set, which are used for classification by SVM.2. Another novel data compression and classification method based on a deep learning network named autoencoder is present in this part. The method includes deep-learning and neural-network related theories. Features in new space computed by deep-learning methods are usually more valuable than that in original feature space. The proposed neural network classifier can utilize the converted data points with deep features to make comparison with existed training set. Only those key data with low similarity can be selected and added into training set. Then SVM classifier uses the labelled data to train the model and test the others.3. This paper proposes a classification method based on CNN theory, which is used for hyperspectral image. This algorithm utilizes multi-layer CNN to learn the features of training set of hyperspectral image, then computes the network structure according to the labels of them. At last, the other data points are classified under the obtained model.