Which metrics to assess and compare the quality of image fusion

Since a little bit more than a decade, a lot of fusion methods for panchromatic and multispectral (P/XS) satellite images have been suggested. But generally, the Quality Assessment (QA) of the outputs is too little taken into account. Sometimes, they are submitted to the visual analysis of interpreters regarding several thematic applications (Subjective QA SQA). Or several quantitative indices (Objective QA OQA) are computed on the basis of a reference image if it exists. Sometimes both approaches are used and compared. So, image QA implies the use of numerous and various criteria that concern as well qualitative ordinal scales as quantitative ones. We define and discuss the concepts of image quality and image QA. The correlation between SQA and OQA is generally poor. This is explained by the fact that OQA metrics don’t take in consideration the characteristics of quality perception by human visual system. But according to our expertise of interpreter it is also important to note that this perception is also influenced by the final use of the images. This use must be clearly defined in order to focus the QA on the quality aspects that regard it. Furthermore, the way of aggregating the various aspects or criteria of QA must also be re-examined. The SQA is performed using qualitative criteria that must be as clearly defined as possible to allow a ranking of fusion methods. Their evaluation scale and relative weight should be discussed regarding all the qualitative criteria and potential uses of the images to allow a global qualitative ranking. The quantitative metrics of OQA are defined and discussed regarding different aspects of image quality: textural, spectral and geometric quality of polygenic images. The spatial scale of these metrics is taken in consideration: global, local or “punctual” QA. Furthermore, their spectral dimensions are also analyzed: monospectral vs multispectral. We suggest a method that illustrates these aspects of OQA on the basis of some fusion results obtained from simulated PLEIADES-HR data in different landcover contexts. This is a pairwise comparison method based on Multi-Criteria Evaluation (MCE) principles: a Multi-Criteria Quality Index (MCQI) is defined. Several quantitative aspects that integrate the characteristics of human vision system are aggregated to answer to the following questions: which method works better, where does it work better and how better does it work ?