Performance Evaluation of Multiple Pan-Sharpening Techniques on NDVI: A Statistical Framework

Abstract

Pan-sharpening is a pixel-level image fusion process whereby a lower-spatial-resolution multispectral image is merged with a higher-spatial-resolution panchromatic one. One of the drawbacks of this process is that it may introduce spectral or radiometric distortion. The degree to which distortion is introduced is dependent on the imaging sensor, the pan-sharpening algorithm employed, and the context of the scene analyzed. Studies that evaluate the quality of pan-sharpening algorithms often fail to account for changes in geographic context and are agnostic to any specific applications of an end user. This research proposes an evaluation framework to assess the effects of six widely used pan-sharpening algorithms on normalized difference vegetation index (NDVI) calculation in five contextually diverse geographic locations. Output image quality is assessed by comparing the empirical cumulative density function of NDVI values that are calculated by using pre-sharpened and sharpened imagery. The premise is that an effective algorithm will generate a sharpened multispectral image with a cumulative NDVI distribution that is similar to the pre-sharpened image. Research results revealed that, generally, the Gram–Schmidt algorithm introduces a significant degree of spectral distortion regardless of sensor and spatial context. In addition, higher-spatial-resolution imagery is more susceptible to spectral distortions upon pan-sharpening. Furthermore, variability in cumulative density of spectral information in fused images justifies the application of an analytical framework to assist users in selecting the most effective methods for their intended application.