Dimension reduction in hyperspectral images

Besides making life easier in many areas, developed digital imaging technology has become a necessity. By the development of technology, imaging technology has also forwarded and advanced imaging opportunities, like hyperspectral imaging, have occured. By the use of hyperspectral imaging, while large sized data with narrow spectral bands has been processed and the success of classification has increased, calculation costs has started to increase. The result was the need to dimension reduction. In this study, different dimensionality reduction methods have been applied to hyperspectral data sets and the classification performance on different classifiers is analyzed. Using PCA and Sparse PCA methods, feature extraction or band selection was applied to reduce dimension and then classified with different types of classifiers. K-nn, SVM, K-means and hierarchical clustering methods has been chosen as learning methods. In this study, the bands are grouped in order to rehabilitate of the results of the Sparse PCA. Unlike classical SPCA method, local maximum points where coefficient of dicsreteness peaked have been observed. Those points were chosen from band groups that were grouped specific range. Band selection process was applied for 3, 5 and 10 bands. Confusion matrix was used to measure the results of supervised classifiers. Accuracy calculation was implemented for the whole system. Also the success of each class was calculated by calculating F1_ scores. Adjusted Rand Index method was used for the performance of measurement of unsupervised classifiers. All dimension reduction and classification processes has been applied to AVIRIS Indian Pines and KSC data sets. It was seen that grouping approach to Sparse PCA has yield successful results. Key words: Hyperspectral images, Dimension Reduction, Band Selection, Sparse PCA