Self Organizing Map based Land Cover Clustering for Decision-Level Jaccard Index and Block Activity based Pan-Sharpened Images

Abstract

Pan-sharpening is very often employed in remote sensing to transform low-resolution multispectral (LMS) images into equivalent high-resolution multispectral images (HMS). Images resulting from pan-sharpening are sharper and more detailed that is resulted by improving spatial features of the multispectral image. One such approach of jointly processing LMS and Panchromatic images is discussed in the present study. The decision-level fusion suggested here involves choosing or combining details from numerous sources by taking decisions while analyzing features recovered from the input images. The proposed methodology is an amalgamation of principal component analysis used for separating spatial and spectral details of LMS, non-subsampled contourlet transform for feature analysis, and Jaccard similarity index and block activity measurement for localized decision-based fusion. The algorithm tries to provide an adaptive approach to address the trade-off between spectral and spatial resolution. Self-Organizing Maps based clustering technique is employed with the intension of grouping the image pixels into three categories soil, water and vegetation. The paper highlights the performance comparison of proposed method with various pixel-level fusion techniques ranging from techniques from Intensity, Hue and Saturation (IHS) transform to Neural Networks based pan-sharpening methods. This comparison is implemented using various reference and non-reference indicators along with Kolmogorov–Smirnov test. Additional analysis using Kolmogorov–Smirnov test is done to statistically analyze spectral degradation. The comparative analysis provides enough evidence that the suggested method yields fused images with enhanced edge details without forgoing the spectral features which was also evident from the mutual information obtained from clustered images. The resulting sharpened images tend to possess good spatial and spectral details that would simplify the automatic image analysis.