Fixed Point ICA Based Approach for Maximizing the Non-gaussianity in Remote Sensing Image Classification

Abstract

To maximize the non-gaussianity in sphered satellite data, many authors have proposed different independent component analysis (ICA) based approaches to classify images by reducing the mixing effect in classes. In multispectral data, few heterogeneous classes have little variation in spectral resolution. Even though, a classified image should exhibit high spectral variance among different classes, while it should be less within a particular class. To improve the classification accuracy in the presence of mixed classes i.e., having similar spectral characteristics, a novel method improved fixed point independent component analysis (IFPICA) is proposed. This method segregates the objects from mixed classes on maximizing the approximation of negentropy, which reduces the effect of quite similar spectral characteristics among different classes. It can easily estimate the independent component of this non-gaussian distribution of data with the help of nonlinearity. Therefore, this nonlinearity helps to optimize the performance of this approach, which minimizes the variance among similar classes. Due to the presence of neural algorithms, it is quite robust, computationally simple and has very fast convergence, in respect to the spectral distributions of satellite images. Hence, this proposed IFPICA approach plays a major role in the classification of satellite images such as road, vegetation, buildings and grassland area. The images used in the study doesn’t have any initial or additive noise, which would obstruct the process of IFPICA algorithm used in the work, therefore preprocessing is not required for noise suppression in this work. The post-processing, e.g., deflation, denoising, filtering, etc. are also not required due to similar reason.