Image fusion methods based on a linear mixing model of multispectral remote sensing data

Kurzfassung

Model based analysis or explicit definition/listing of all models or assumptions used in the derivation of an image fusion method allows us to understand the rationale or properties of existing methods and shows a way for a proper usage or proposal/selection of new methods ‘better’ satisfying the needs of a particular application. Most existing pan-sharpening methods are based mainly on the two models or assumptions: spectral consistency for high resolution multispectral data (physical relationship between multispectral and panchromatic data in a high resolution scale) and spatial consistency for multispectral data (so-called Wald’s protocol first property or relationship between multispectral data in different resolution scales). Additionally, it can be seen/shown easily that the following two popular groups of methods: spectral transformation (e.g. Intensity-Hue-Saturation (HIS), Principal Component Analysis (PCA) and Gram–Schmidt orthogonalization (GS)) and filtering (e.g. High Pass Filtering (HPF) and Multi-Resolution Analysis (MRA)) based methods are based on a pure pixels assumption. Thus, their usage for mixed pixels (quite common situation in real remote sensing applications) can lead to wrong image fusion results. Very few methods exist which can treat mixed pixels in a correct way. Two methods, one based on a linear unmixing model and another one based on spatial unmixing, are described/proposed/modified which respect models assumed and thus can produce correct or physically justified fusion results. Earlier mentioned property ‘better’ should be measurable quantitatively, e.g. by means of so-called quality measures. The difficulty of a quality assessment task in multi-resolution image fusion or pan-sharpening is that a reference image is missing. Existing measures or so-called protocols are still not satisfactory because quite often the rationale or assumptions used are not fulfilled. From a model based view it follows naturally that a quality assessment measure can be defined as a combination of error model residuals using common or general models assumed in all/most fusion methods. Thus in this paper a comparison of the two earlier proposed/modified unmixing based pan-sharpening methods together with some already existing model based methods and several other popular methods is performed. Experimental validation/verification is carried out in the urban area of Munich city for optical remote sensing multispectral data and panchromatic imagery of the WorldView-2 satellite sensor. The quality assessment of image fusion or pan-sharpening results is performed using a newly proposed measures based on common or general model error residuals and their combinations. Preliminary results confirm ideas of the author and show a great potential for future applications such as clustering, classification, matching and change detection.