Comparison of Support Vector Machine, Artificial Neural Networks and Spectral Angle Mapper Classifiers on Fused Hyperspectral Data for Improved LULC Classification

The present paper focuses on analyzing the performance of three different advanced classifiers on hyperspectral data for improved land use/land cover classification. Considering the limitations of spatial and spectral resolution of high resolution multispectral and hyperspectral datasets, image fusion techniques were used for obtaining a spectrally and spatially rich space borne hyperspectral data. The Gram Schmidt sharpening, Principal Component sharpening and Color Normalized spectral sharpening were used for fusing the 30m, 242 band EO-1 Hyperion image with the 1.8m eight band World View 2 data. Correlation, SNR and entropy measures were adopted for analyzing the performance of the fused outputs. Color Normalized spectral sharpening was observed to exhibit highest correlation of 0.78 between spectra of Hyperion and fused images. Classification of the fused and unfused images was performed using three different classifiers - Spectral Angle Mapper, Artificial Neural Networks and Support Vector Machine by taking end members from the image and the ground spectra. Confusion matrix was generated using a ground truth points obtained during field visit. Support Vector Machine classifier with RBF kernel and lOO penalty value was observed to give an improved classified output than the remaining classifiers with an accuracy of 89.56% and a kappa coefficient of 0.88%.