RS data fusion by local mean and variance matching algorithms: their respective efficiency in a complex urban context

Two methods of remote sensing (RS) data fusion using local mean and variance matching (LMVM) algorithms are tested on one complex urban subset of an IKONOS image of Toulouse. This image is composed by 4 multi-spectral (XS) bands and 1 panchromatic (P) band. Their spatial resolutions are respectively 4 and 1 m. The first scheme is a band by band fusion and the second one is a combined bands fusion using the INR (intensity normalized ratio) transform. Both methods depend on one single parameter, the size of a convolution window used for local means and variances computation. In an urban area, the setting of the window-size parameter is very highly sensitive to the structures with high spatial frequency and to the variable urbanization rate. As we have demonstrated in the past, the window-size determines the importance of the textural information coming from the panchromatic high spatial resolution (P HSR) image injected in the fused product. The weight of injected textural information is generally proportional to the magnitude of this parameter but the spectral information conserved from the multi-spectral low spatial resolution (XS LSR) image decreases with this parameter. The respective efficiency of the two schemes in the /sub c/omplex urban context of Toulouse is compared for different window-sizes using global correlation and visual computer assisted image interpretation. From this comparison, the 11/spl times/11 convolution widow-size gives very good results for the first scheme on spectral and textural points of view. The second scheme is not appropriate to get high quality products.