Accuracy improvement of high resolution satellite image georeferencing using an optimized line-based rational function model

Abstract

ABSTRACTThe georeferencing procedure of high resolution satellite images (HRSIs) using proper mathematical models is an important step in 3D spatial information extraction. Since the 2000s, line-based mathematical models are considered more in photogrammetric communities. This could be due to the unique characteristics of linear features such as more reliable procedure of automatic matching as well as abundance of linear features in satellite images. In addition, the irreplaceable characteristics of Rational Function Models (RFM) such as generality and its independence to the sensor model makes it a proper mathematical model for this purpose. Although, traditional selection of the best order and terms’ combination of RFM due to the deficiency of physical interpretation of terms is not easily feasible. Hence, in this article, an optimization algorithm based on a binary particle swarm optimization (PSO) is developed to determine the optimum uncorrelated terms’ combination of a line-based RFM. For this purpose, a population of particles (representative of different RFM structures) is initialized randomly with a string of binary values, indicating the presence or omission of the corresponding terms. This is followed by directly using some conjugate lines in the image and object spaces as ground control lines (GCLs) to solve the unknown parameters of the RFM for each particle. The root mean square error (RMSE) of some check points (CPs) for each particle is considered to be its cost function and used to update the velocity of particles. The procedure is repeated to reach a stopping criteria. A comprehensive evaluation on the proposed model in comparison to the traditional line-based RFM is examined. Two different HRSIs (a GeoEye and an Ikonos Images) over different areas of Iran are used for this purpose. Based on the results, a traditional line-based RFM could not reach to sub-pixel accuracy. In contrast, the results show the potential of the proposed optimized line-based RFM to increase the accuracy to better than 0.8 pixel as well as reduce the systematic errors and the number of required control information, significantly.