Analogue-based colorization of remote sensing images using textural information

Abstract

Abstract Satellite images are richer than ever before. For example, new Landsat-8 images with their 11 bands carry much more information than older generations of satellites. These differences in spectral representation imply a major difficulty for assessing long-term land surface changes. The easiest solution is to reduce the information of the most recent product, for example by only keeping a subset of the Landsat-8 bands that matches old imagery. To avoid such loss of information, we propose a new method based on multiband spatial pattern matching. We are focusing on increasing the spectral resolution of archive satellite images to the same level of spectral resolution and coverage as modern imagery. Our method uses analogous scenes taken from modern satellites, which have conceptually the same role as the training images used in multiple-point geostatistics simulation. The spectral characteristics of the training image are then transferred to a target archive image, where new synthetic spectral bands are generated. A spatial pattern matching procedure is used to control this transfer, resulting in preservation of spatial and spectral coherence in the results. We illustrate the methodology on Landsat 8 and Corona imagery. The proposed method was benchmarked against other state-of-the-art colorization techniques, and it shows globally better results.