A Method for Change Detection with Multi-temporal Satellite Images Using the Rx Algorithm

The use of new, high-resolution spectral sensors, in contrast with low-resolution, increases both the amount of information acquired about land cover at local scales, and the geometric detail and accuracy. However, the algorithms needed for high-resolution image processing are more complex than those needed for their low-resolution counterparts. This paper's main objective is to present a new method for change detection for bi-temporal, multi-spectral, high-resolution satellite imagery. The method focuses on detecting anomalies in two images with the RX algorithm, and then analysing the differences. As a case study, the method is tested with two SPOT5 satellite images pansharpened to a resolution of 2.5 m. Most of the changes that happened to manmade objects between the two dates are obtained by this method. * Corresponding author. 1. INTROUCTION Many algorithms and techniques have been proposed in the last two decades for change detection using multi-temporal satellite data; the most popular techniques are based on algebraic operations, transformations, classification, neural networks, etc.; see a review by Lu et al. (2004). Low-resolution satellite imagery commonly uses pixel-by-pixel change detection methods. This cannot be applied to high-resolution satellite imagery (i.e. SPOT5), due to the complexity of the scene; these images have greater detail, with many new signatures from different materials appearing in high-resolution images that were not detected in low-resolution images. For example, shadows present a problem in high-resolution imagery but not low-resolution imagery. We discuss SPOT5 imagery here, but other satellites have even higher resolution, and the change detection method presented here could be applied with the proper modifications. These satellites include Ikonos, Quickbird, WorldVie-1 and many more to be launched in the near future; among them, the WorldView-2, scheduled for launch in 2009, will provide eight-band multispectral imagery for mapping and monitoring applications. It will offer a ground resolution of 0.5 m panchromatic and 1.6 m multispectral. These high-resolution satellite images offer significant cost savings compared to aerial photography as a result of the larger footprints, which means that less ground control and less processing are necessary for orthorectification. High-resolution satellites have more frequent revisit times than aerial surveys, and therefore there is the potential for automatic feature change detection. The gap between aerial photography and satellite imagery is progressively being bridged. One proof of this is national map agencies show increasing interest in this type of imagery; they are acquiring it frequently. This imagery is very useful for change detection of manmade objects, and especially in suburban and urban areas, but most of this work is being done manually. Data from high-resolution sensors clearly offer exciting new challenges and opportunities for researching semiautomatic techniques that could help people working in the cartographic industry. 2. DATA CHARACTERISTICS AND PREPROCESSING To see the method's potential, two images from the SPOT5 satellite have been used. The SPOT5 satellite was launched in 2002 and captures panchromatic images with a resolution of 2.5 m, and multi-spectral images with a resolution of 10 m. The first image was taken at 9:30 a.m. on 7-24-05 and the second at 11:20 a.m. on 8-12-06; both are of the same area southeast of Madrid, Spain, on a mostly flat terrain. Azimuth and elevation of the sun were (138.03, 64.97) and (150.07, 62.37), respectively. Even though both images were taken in the summer and with similar azimuth and elevation of the sun, we applied a radiometric normalisation to both images in order to obtain the closest possible brightness conditions between the images. The normalisation we have applied is called multivariate alteration detection (Canty, 2007). We took 20 control points (with less than 1 m RMS) uniformly distributed throughout the study area. A subset of the control points were chosen and were subsequently used as independent checkpoints. These were also distributed approximately uniformly. With the metadata from the SPOT5 satellite, the control points and a DEM (5 m resolution) acquired from the Spanish National Map Agency (Instituto Gegráfico Nacional, IGN), we orthorectified both images. Then we applied PCA pansharpening to both images in order to obtain two images with four multispectral bands, with 2.5 m resolution each. Then both images were co-registered until errors were under one pixel.