Hyperspectral Image Segmentation Using Evolutionary Multifactorial Spectral Analysis for OMEGA Dataset

Abstract

Mars region is being imaged with an exceptional combination of spectral and spatial resolution spectrometer using OMEGA instrument. The hyperspectral images of Mars provide spectral range and chemical species with high resolution. This paper presents a novel unsupervised segmentation algorithm named as evolutionary component analysis for remotely sensed hyperspectral image data for material identification in the spatial and spectral information. Sparse multinomial logistic regression (SMLR) algorithm is initially employed to learn the posterior probability distributions from the spatial and spectral information of the images containing class imbalance information to infer the class distribution of the testing hyperspectral data. Evolutionary multifactorial spectral analysis (ESA) helps to characterize noise and extremely mixed pixels with less training set with high training quality and utility with respect to spectral signatures and its spectral changes with less interaction for classification and end-member detection. The proposed segmentation approach based on ESA is investigated and estimated using both real and simulated hyperspectral datasets. ESA is evaluated for the endmember extraction in the mixed pixel revealing up-to-date performance when compared with advanced hyperspectral image classification techniques. The combined spatial–contextual information (ESA + SMLR) characterizes a state-of-the-art contribution in the research field of material identification. The proposed approach is exposed to present proper classification of the minerals of Mars surface in both the spatial and the spectral domain in short span of time.