Hyperspectral Image Feature Extraction And Classification Based On Autoencoders

By combining spectroscopy and imaging technology, hyperspectral imagery canobtain continuous data on both spatial and spectral dimensions simultaneously. Inrecent years, the spectral and the spatial resolution of hyperspectral images has beenimproving rapidly. However, besides the plenty of information this improvement onspectral resolution brings, it has also incurred new pressure and challenges towardpost-processing algorithms, such as classification. Feature extraction, as a methoddealing with the high dimension problem, has become increasingly important. Thenonlinear component underlying the high-dimensional data are increasinglyimportant thus cannot be ignored anymore. On the other hand, the growing spatialresolution increases the correlation between pixels, which renders spatialinformation-based classification possible.The paper discusses and gives a clear definition of the "nonlinear" feature inhyperspectral images, and explores the sources of the nonlinearity of hyperspectralimagery by analyzing its imaging process. Through direct observation from thefeature space of hyperspectral data, and by using statistical hypothesis testingmethods, the existence of nonlinear characteristics has been verified. For nonlinearfeature extraction, this paper carries research on one of the deep learning model-autoencoder. Focusing on autoencoder and its deep structure, stacked autoencoder,we study the structure of these models, their training methods and visualizationanalysis.Oriented by the application of hyperspectral data classification, this paper usesthe stacked autoencoder to extract deep features for classification. We proposed2schemes–Autoencoder-Support Vector Machine (AE-SVM) and StackedAutoencoder-Logistic Regression (SAE-LR). The appropriate feature numbers anddepths are discussed on these two classification schemes. Experiments show thatautoencoder as a feature extractor can provide better features than traditional linearfeature extraction methods; and SAE-LR outperforms support vector machine on avariety of result evaluation measurements.Although as a feature extractor, autoencoder have obvious advantages comparedwith traditional methods, it may still be possible to further improve accuracy byintroducing spatial information. To utilize spatial information, this paper proposes amethod which uses principal components of pixel neighborhood, and further integrateit with spectral features by using SAE-LR to form a joint spectral-spatialclassification framework. Experiments show that this classification framework canfurther increase the classification accuracy, thus pushes the classification result beyond support vector machine and the aforementioned spectral information basedclassification.