Evaluation of Wavelet Transform Algorithms for Multi-Resolution Image Fusion Saim Muhammad, Monica Wachowicz and Luis. M. T. de Carvalho Center for Geo-Information Wageningen UR, PO Box 47 6700 AA Wageningen, The Netherlands Telephone: +31.317.474764 – Fax: +31.317.474567 http://www.geo-informatie.nl/ E-mails: muhammad.saim@student.girs.wau.nl m.wachowicz@alterra.wag-ur.nl , luis.carvalho@staff.girs.wau.nl Abstract Wavelet Transforms can be used for multi-resolution image fusion at pixel level, as they work both in spatial and spectral domains and result in the preservation of spatial of spectral details of input images. Different wavelet transform algorithms have been developed and applied to a variety of applications such as noise suppression, filtering, image restoration, image compression, and astronomical applications. This paper explores the use of current developed wavelet transform algorithms for multi-resolution fusion of satellite images. The aim is to investigate how appropriate these wavelet transform algorithms are for this multi-resolution image fusion. Five different types of wavelet transform algorithms are selected and the results are evaluated by comparing their spatial and spectral quality with the spatial and spectral qualities of their source satellite images (i.e. Ikonos Panchromatic and Multispectral, and Landsat TM). The findings show that different wavelet transform algorithms have a "preservation tradeoff" between the spatial quality and spectral quality. Due to the frequency shift limitation of wavelet transform, it can preserve the spatial and/or spectral details of the input images for a certain number of levels. Keywords: multi-resolution image fusion, wavelet transform, spatial quality, spectral quality 1 Introduction A vast literature is available and provides a system-level discussion about multi-sensor fusion technologies as they relate to command, controls, communication, and intelligence [1], [2] and [3]. The technological aspects have been considered within detection (or decision) theory, estimation theory, digital signal processing, and parametric and non-parametric data fusion techniques including fuzzy logic, neural networks, and voting logic [4] and [5]. However, it is only recently that data fusion techniques have been considered for remote sensing applications to improve the spatial resolution of multi-spectral images [6] and their classification performance [7]. This reality coincides with an exponential increase in digital data generated by Earth Observation Systems. A full set of sweeps from NEXRAD-88D radar can record 13 megabytes of data over a five-minute period over 120 radar sites across the U.S. The Moderate Resolution Imaging Spectroradiometer (MODIS) developed for global remote sensing of clouds, aerosols, water vapor, land, and ocean properties provides 1.29 Gbytes per hour. With so much data comes the problem of combining (fusing) these data and extracting useful information from them. Most of the models that have been proposed for data fusion consist of three levels. These levels are the signal level (also called pixel, measurement), the feature level (attribute), and the decision level. They have been described in such a way that we begin with the pre- processing of data, followed by feature extraction. Having extracted the features, object recognition (also called identification) is performed by statistical techniques, or geometric models. The results are usually partitioned into groups that represent objects belonging to the same class. Selection of the level for multi-resolution image fusion depends on the characteristics of the input images and the desired output. In this paper, we focus on multi- resolution image fusion at the pixel level. A multi-resolution fusion model is proposed based on both spectral and spatial characteristics of satellite images (raw data). Several approaches have been proposed for the implementation of multi-resolution fusion models at the pixel level. Some examples are linear superposition [8], optimization methods [9], image pyramids [10] and wavelet transform [6] and [9]. However, we see a substantial 1573 ISIF © 2002