Intermodality models in pan-sharpening: analysis based on remote sensing physics

Abstract

Many intermodality models (IMMs) have been established to define how panchromatic (Pan) high-frequency details should be injected into the multispectral (MS) bands in pan-sharpening. In this paper, IMMs are categorized into two types, namely regression coefficient (RC) models and intensity modulation (IM) models. The RC is simply the ratio of the local covariance of x and y to the local variance of x, where x and y are the approximation of the Pan image at the MS scale and the MS image, respectively. To compare these two IMM types, an ideal IMM representing the proper amount of high-frequency details which should be injected is formulated. This ideal IMM is exactly the same formulation as the RC model except that x and y are high-resolution Pan and MS images, respectively. Therefore, compared to the ideal IMM, the RC model is influenced by the scale effect since the local variance and covariance change as functions of scale. The IM model, represented by the model in the high-pass filtering (HPM) method, may introduce improper injections which are affected by landscape complexity, spectral response function, atmospheric conditions, and offset values for converting radiance into digital number (DN). Using the information from the high-resolution Pan image, an improved HPM method is proposed for shadow regions, which are quite common in very high-resolution images. Experimental results using a WorldView-2 image covering an urban area and a QuickBird image covering a rural area confirmed our analytical conclusions on different IMMs and indicate that the improved HPM model performs better than the original HPM.