Evaluation of Two Applications of Spectral Mixing Models to Image Fusion

Abstract

Many applications in remote sensing require fusing low-resolution multispectral or hyperspectral images with high-resolution panchromatic images to identify and map materials at high resolution. A number of methods are currently used to produce such hybrid imagery. Until now, these methods have only been evaluated independently, and have not been compared to one another to determine an optimum approach. Two different operations are involved in creating high-resolution material maps. One task is to unmix hyperspectral images into the constituent materials (endmembers). The other task is to fuse low-resolution imagery containing spectral detail with high-resolution image(s) containing fine spatial information. This research performs a quantitative test of three image fusion procedures. The first method is to fuse, then unmix. One sharpens low-resolution multispectral data using a panchromatic image, producing a set of high-resolution multispectral images. These images are then separated into a series of high-resolution endmember maps, locating the materials within the scene. The second method is to unmix, then fuse. In this approach, one first separates the low-resolution multispectral data into a series of material maps using a recently developed adaptive unmixing algorithm. These maps are fused with the panchromatic image in a stage called sharpening, to produce high-resolution material maps. The final method is also an unmix-then-fuse approach. Here, the low-resolution material maps are created using traditional image-wide spectral unmixing methods. The resulting images are fused with the panchromatic image to produce sharpened material maps. In this paper, the three image fusion procedures described above are evaluated for their radiometric and unmixing accuracy. The results show that the adaptive unmixing algorithm is superior to the traditional imagewide methods. Differences between quantitative and visual evaluations indicate an improved error metric must be developed.