Research On Semi-supervised Classification Algorithms For Hyperspectral Imagery

With the rapid development of remote sensing technology, the image provided by imaging spectrometer has been widely researched and applied into many fields. Hyperspectral imagery classification is a basic content of hyperspectral data analysis and processing. A difficulty in data processing and information analysis is that there often exists limited number of labeled samples compared to the high-dimensional dataset, which may lead to the Hughes phenomenon in supervised classification. And due to the high cost of acquiring labeled training samples, traditional supervised classification methods show great limitations.Semi-supervised classification learning, which attempts to use the labeled samples together with large amount of unlabeled samples, is of great interest both in theory and in practice.Semi-supervised classification is to make full use of spectral and spatial information that the unlabeled data carries and to train classifiers with good generalization performance and high classification accuracy. Tri-training and self-training are two commonly used semi-supervised classification methods. For tri-training method, there is no obvious diversity between the three classifiers with small training samples, which restricts the improvement of classification accuracy. As for self-training,it's easy to produce mislabeled samples, which will decrease the classification precision when incorporated into the training set.In this paper,we research on semi-supervised classification algorithms for hyperspectral imagery at the basis of previous studies. The main contents are as follows:1. Tri-training method based on active learning (AL) and differential evolution (DE)algorithm. In the proposed framework, AL was used to select the most informative unlabeled samples, with which new samples were generated by DE algorithm before training each of the three classifiers. The newly generated samples will be labeled and added to training set to train the base classifiers for tri-training. The proposed algorithm was experimentally validated on real hyperspectral datasets, indicating that the proposed framework can utilize the unlabeled data effectively and achieve high accuracy and efficiency compared with state-of-the-art algorithms when small labeled data is available.2. Semi-supervised classification algorithm based on spatial-spectral clustering. In the proposed framework, spatial information extracted by Gabor filter was stacked with spectral information first. After that, AL was used to select some most informative unlabeled samples as the candidates which may be added into the training set later. Self-training was employed to predict the labels of the unlabeled samples. And we combined the result of probability model-based SVM (Pro-SVM) with the output obtained by clustering technique to select samples with high reliability and to remove samples that were erroneously labeled in the self-labeled process. Moreover, spatial information was incorporated into the clustering technique to improve classification accuracy. Meanwhile, the probabilistic output of SVM can provide the probability of any example belonging to a class. In the labeling process, we set a probability threshold to combine the Pro-SVM and the clustering technique. Two hyperspectral images were used for experiment with the proposed algorithm, which proved that this algorithm can label unlabeled samples effectively and have high classification accuracy and Kappa coefficient.In the proposed algorithms, the using of AL algorithm reduced the labeling cost for unlabeled samples effectively and improved the running efficiency. The using of DE algorithm can fulfill the labeled data distribution and introduce more diversity to multiple classifiers. The clustering technique with spatial and spectral information improved the accuracy of clustering. And the using of Pro-SVM made the labeling process more reasonable especially for mixed pixels.