Pan-Sharpening of Very High-Resolution Multispectral Images via Generalised Laplacian Pyramid Fusion

Cet article presente une solution au probleme de fusion entre des images multispectrales et des images panchromatique a haute-resolution. La methode repose sur la pyramide Laplacienne generalisee, qui est obtenue par une soustraction recursive entre l'image et Ie resultat d'un filtrage passe-bas. Elle fonctionne meme pour des rapports de resolution non en tiers entre les images fusionnees. Contrairement a d'autres processus de fusion d'images, la decomposition proposee lors de I'analyse multiresolulion n'est pas sous~echantillonnee de fac;on critique, afln d'eviter les problemes d'imprecision et de degradation qui pourraient apparaitre dans I'image fusionnee. En effet. ces defauts proviennent souvent d'informations supprimees dans les termes d'aliasing. La methode de fusion substitue de fac;on selective les hautes frequences spatiales de I'image haute resolution (P) acelles des images aplus basse resolution spatiale (MS). Plusieurs modeles d'injection des details, globaux ou locaux sont presentes. La distorsion spectraJe du produit fusionne est abordee grace a un modele d'injection capable de rorcer la similitude des distributions entre les bandes MS reéchanlillonnees (prises com me reference spectrale) et Ie produit fusionne en chaque pixel. Les rasultats sont presentes et discutes sur les donnees provenant du capleur aerien hyperspectral a Ires haute resolution MIVIS et simulant des donnees SPOT 5 sur une zone urbaine

This work presents a viable solution to the problem of fusion of multispectral (MS) images with high-resolution panchromatic (P) observations. The method relies on the generalized Laplacian pyramid (GLP), which is obtained by recursively subtracting from an image its low-pass version, and works even with fractional scale ratios between the data to be merged. Unlike other multiscale fusion schemes, the decomposition proposed for the multiresolution analysis underlying the fusion procedure is not critically sub-sampled, thus avoiding possible inaccuracies and impairments in the fused images, originated from the missing cancellation of aliasing terms. The fusion method selectively performs spatial-frequencies spectrum substitution from the higher-resolution (P) image to the coarser (MS) bands. Several detail injection models, both global and local, are presented. Spectral distortion of the fused product is addressed in devising an injection model that is capable to constrain to zero the spectral angle between resampled MS bands, taken as reference of spectral fidelity, and fused product at every pixel. Results are presented and discussed on very high-resolution SPOT 5 data of an urban area, simulated by means of hyperspectral data collected by the MIVIS air-borne spectrometer.