Image Fusion Using Cokriging

Scientists in various disciplines are faced with huge amounts of data that need to be studied and analyzed. NASA alone has around 18 satellites with over 80 sensors, all of which send a tremendous amount of data from around the globe continuously. An important step in modern data processing applications where data are gathered from multiple sources is data fusion. Data fusion is defined as the process of dealing with information from multiple sources to achieve refined and improved information for decision making [1]. Image fusion is a subset of the general data fusion problem where data being fused are images. The goal of performing image fusion is usually to increase either the spatial or spectral resolution of images involved. One particular case of image fusion is pan-sharpening. Pansharpening is a technique which deals with the limitations of sensors in capturing high resolution multispectal (MS) images [2]. That is panchromatic (Pan) images have high spatial resolution and low spectral resolution. On the other hand, MS images have high spectral resolution, since they cover a narrower wavelength range, but have a lower spatial resolution. Image fusion is then used as a tool to create a high spatial and spectral resolution image given Pan and MS images. In this paper we show how to apply fusion for the purpose of pan-sharpening multispectral Landsat ETM bands by using cokriging. We employ the cokriging interpolation method for image fusion of remotely sensed data [3], [4]. In particular, we show preliminary results on applying a variant called ordinary cokriging for pan-sharpening of multispectral images from the Landsat 7 sensor. We initially proposed cokriging for image fusion in [5] and showed preliminary results on increasing the spectral resolution of ALI using Hyperion. In this paper, we address the problem of increasing the spatial resolution of multispectral bands of a sensor using a panchromatic image through. We then evaluate both spectral and spatial quality of our fused images through few quantitative measures. We also compare our results to those obtained from more traditional approaches based on principal component analysis and wavelets.