A Topographical Feature Extraction Approach for Classification of Soil Hyperspectral Image

Abstract

Hyperspectral images, having more than hundreds of bands and very high spectral resolution, endeavors a favorable approach for classification of the soil. The accuracy and viability of visible-near infrared (Vis-NIR) hyperspectral imaging proved to be more powerful, as these images have both spatial and spectral information. The purpose of this study is to perform classification on AVIRIS hyperspectral images based on their geographical nature of the soil with endmember selection while comparing various classification models. The classification techniques of these hyperspectral images were analyzed using small fractions among the number of training samples and their spectral features. The study analyzed that use of Constrained Energy Minimization technique yields better results among the various supervised classification techniques. Also, when the hyperspectral data has to be classified using unsupervised learning techniques like K-Means and ISODATA, K-Means performed better than ISODATA with the accuracy of 98.3%.