Pixel Level Fusion of Panchromatic and Multispectral Images Based on Correspondence Analysis

A pixel level data fusion approach based on correspondence analysis (CA) is introduced for high spatial and spectral resolution satellite data. Principal component analysis (PCA) is a well-known multivariate data analysis and fusion technique in the remote sensing community. Related to PCA but a more recent multivariate technique, correspondence analysis, is applied to fuse panchromatic data with multispectral data in order to improve the quality of the final fused image. In the CA-based fusion approach, fusion takes place in the last component as opposed to the first component of the PCA-based approach. This new approach is then quantitatively compared to the PCA fusion approach using Landsat ETM , QuickBird, and two Ikonos (with and without dynamic range adjustment) test imagery. The new approach provided an excellent spectral accuracy when synthesizing images from multispectral and high spatial resolution panchromatic imagery. Introduction Many Earth observing satellites (sensors) are providing increasingly high-spatial resolution multispectral data. However, two major factors limit a remote sensing sensor’s ability to collect high spatial resolution, multi-spectral data (Zhang, 2004). First, the incoming radiation energy to sensor is limited by optics size. Second, the data volume to be collected and stored by the sensor increases exponentially with higher spatial resolutions. Thus, satellites, such as QuickBird and Ikonos, bundle a 4:1 ratio of a high-resolution panchromatic band and lower resolution multi-spectral bands in order to support both color and best spatial resolution, while minimizing on-board data handling needs. The on-ground fusion of panchromatic and multi-spectral bands may provide an improved product to users, dependent upon the ability of the fusion technique to accurately reproduce a synthetic (fused) imagery from a multispectral imagery while improving the spatial resolution. Hence, many fusion techniques are developed to integrate both panchromatic and multispectral data in order to increase the spatial resolution of the former (e.g., Cliche et al., 1985; Price, 1987; Welch and Ehlers, 1987; Chavez et al., 1991; Ehlers, 1991; Shettigara, 1992; Yesou et al., 1993; Zhou et al., 1998; Liu and Moore, 1998; Zhang, 1999; Lemeshewsky, 1999; Ranchin and Wald, 2000; Laben, and Brower, 2000; Ranchin et al., 2003; Cakir and Khorram, 2003; Chen et al., 2005) Pixel Level Fusion of Panchromatic and Multispectral Images Based on Correspondence Analysis Halil I. Cakir and Siamak Khorram A critical consideration is how to integrate spatial information present in the panchromatic image but missing from the low-resolution multispectral data. Many techniques transform multispectral data from color space to a new space in order to have at least one component highly similar to panchromatic data such as the principal component analysis (PCA) or intensity-hue-saturation (IHS)-based techniques. By substituting this component with panchromatic data and then performing inverse transformation to original color space, a new multispectral image with the spatial resolution of panchromatic data is achieved. However, accurate production of the synthetic image is dependent upon the spectral equality of the substituted component and the panchromatic band (Švab and OŠtir, 2006). Thus, panchromatic data is preprocessed (i.e., histogram matched) before the substitution in order to increase the similarity to the substituted component. In general, transformation methods that result in a component more similar to the panchromatic band do better in terms of the spectral accuracy of the synthesized images. Some recent studies have focused on this aspect to improve the fusion process such as the new modified-IHS proposed by Siddiqui (2003) and FFT-enhanced IHS method by Ling et al. (2007). One of the most widely used fusion approaches is based on the principal component analysis of the images. PCA is a multivariate statistical technique that deals with the internal structure of matrices. It breaks down or partitions a resemblance matrix into a set of orthogonal (perpendicular) axes or components. Traditionally, this matrix consists of variance-covariances or correlations (If a correlation matrix is used for principal component calculation, it is also known as “factor analysis” or “standardized principal components.”). Each PCA axis corresponds to an eigenvalue of the matrix. Given an image with n-number of bands, n-number of principal components can be calculated. PCA is useful for image encoding, image data compressing, image enhancement, digital change detection, multitemporal dimensionality, and image fusion (Pohl and Genderen, 1998). Some image fusion applications of the PCA method in the literature are given by Chavez et al. (1991). A more recent multivariate method, correspondence analysis (CA) was developed independently by several authors. An algebraic derivation of CA is often accredited to PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Feb r ua r y 2008 183 Center for Earth Observation, North Carolina State University, Campus Box 7106, Raleigh, NC 27695 ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 74, No. 2, February 2008, pp. 183–192. 0099-1112/08/7402–0000/$3.00/0 © 2008 American Society for Photogrammetry and Remote Sensing M-16.qxd 1/8/08 3:03 PM Page 183