A Comparative Investigation of Image Fusion in the Context of Classification

Image fusion assists in visual interpretation, mapping, change detection and many other applications. Multispectral and Panchromatic images are fused to produce images having enhanced spatial and spectral properties. These properties are generally distorted from original images. The aim of this paper is to identify the effectiveness of the several fusion techniques based on the distortions and applications. This paper employs seven image fusion techniques namely, Brovey transform, intensity hue saturation, high pass filter, principle component analysis, UNB Pansharpening, wavelet transform and multiplicative, available in various commercial image processing software. The data for this study are panchromatic image of Cartosat-1 and multispectral image of IRS - P6 LISS 4 sensor of study area, Bhopal Municipal Corporation area, M.P. State, India. The effectiveness of image fusion techniques is determined by quantitative and qualitative assessments. Quantitative assessment is divided into two parts: 1) assessment of fusion techniques by statistical parameters and 2) accuracy assessment of land use maps generated from the fused images. For part 1, three parameters namely, mean bias, correlation coefficient and Q4 quality index, have been used. Based on the results of part 1, UNB Pansharpening and wavelet transform are the best among seven fusion techniques. For part 2, Gaussian and Artificial Neural Network classifiers have been used to generate land cover maps. However, the accuracy results are inconclusive to identify a single best method. Nevertheless, image fusion by wavelet transform has provided best results in both the sector. Hence, wavelet transform is concluded as the best among selected fusion techniques.