Effects of data transformation and classifier selections on urban feature discrimination using hyperspectral imagery

Hyperspectral remote sensing has been used in various applications which include urban applications. Classifying hyperspectral remote sensing data from urban environments is challenging due to spectrally heterogeneous materials with similar spectral properties. There is a lack of studies on the use of hyperspectral technology in urban mapping in Malaysia although it has widely been used in other countries. The selection of mapping techniques in classification which is the selection of data transformation and classifier selections are very essential to acquire maximum mapping accuracy. This research was conducted to study the effects of data transformation and classifier selections in urban feature discrimination using hyperspectral imagery. Two techniques of data transformation are tested in this study which are the spectral derivative and wavelet transformations. Various wavelet parameters which are the selection of wavelet transformation techniques, mother wavelets, number of vanishing moments and scale or level decompositions have been tested in this study. The selection of classifiers such as Minimum Distance to Mean, Spectral Angle Mapper and Support Vector Machine are also tested in this study. The performance of each parameter tested in this study is assessed through their classification accuracy. McNemar statistical test is used to test the significance difference between two classification results. Three hyperspectral images from two different sensors are tested in this study which are two images came from AisaEAGLE sensor while the other image acquired by AISA CLASSIC sensor. The results show that each transformation parameter and classifier selected gave different results. The classification accuracy derived from derivative transformation is lower than the classification accuracy of reflectance. The right selection of wavelet transformation parameters can give maximum classification accuracy. There is no best wavelet transformation parameters can be determined since the best wavelet transformation parameters of all images are different. Classification using Support Vector Machine gave better accuracy than other classifiers for all images and more robust as it is not affected by the types of data used. The results clearly show the advantages of the Support Vector Machine and wavelet-based data in terms of accuracy. They significantly outperform the other method and achieve overall higher accuracies. Thus, both methods can be considered attractive and useful for the classification of urban hyperspectral data. Wavelet based images of rbio2.2 (Scale-8CWT) of first data set and bior2.8 (Scale-16 CWT) of second data set, and reflectance image of third data set gave highest classification accuracies using SVM classifier which are 96%, 98.4% and 98.4%.

Preview PDF ITMA 2012 13R.pdf Download (920kB) | Preview

Actions (login required)

View Item