Researches On Feature Learning And Classification Of Hyperspectral Images

Recently, remote sensing technology has been developed with the increasing resolution and channels in hyper spectrometer. It is widely used in ecology, geology, hydrological science, agronomy and military application. However, the high dimensional features and large numbers of data bring heavy computational cost, noise distrubation and difficulties on labelling. Therefore, feature reduction and feature extraction are the most popular topics in hyperspectral image processing. This paper mainly discusses algorithms of feature learning and tensor-based classification, by which we can improve the robustness of features and reduce the computational cost, leading to the improvement in hyperspectral classification. The main researches are as follows:(1) A tensor-based semi-supervised scale cut algorithm is proposed as feature learning method for classification. In this work, we utilize newly-built tensor theory, because of its combination of spectral and spatial information within a certain hyperspectral image as a series of 3-dimensional blocks. Then we utilize scale cut method, according to the similarities among various classes and that within same classes, to calculate the mapping function based on SVD. Experiments show that the useful blocks with spectral and spatial information are able to find out an effective feature space and better results of classification via SVM are obtained.(2) A deep-learning based band selection in hyperspectral image processing is present. We utilize autoencoder and stack autoencoder to extract features of each band of the image. The softmax classifier computes the correlation between the bands by measuring possibilities. Spectral clustering is used to label the similar bands into several groups and the centers of groups are selected as the final result with small redundancy At last, the original dataset are modified by reserving the selected bands and make discriminative decisions on it. Experiments show that the selected bands have better results in classification.(3) A multi-task based joint sparse representation classifier is proposed for hyperspectral image classification. In the pre-processing step, samples with similar features are separated as several tasks. Samples from different tasks share different weights in the selection of atoms in joint sparse representation classitier method. By the tramework of multi-task, we can further improve the sparse representation classification algorithm.