Research On Geometric Correction Of Remotely Sensed Imagery Based On EIV Model

Raw images usually contain significant geometry distortion. Corrections on thegeometric distortion for raw images are the prerequisite for multiple source remote sensingdate fusion, integration, processing and analysis. Analysing the reasons of the errors andreducing the errors in the geometric corrections are of importance for users to clearlyunderstand error or uncertainty properties and to increase the correction accuracy.A primary method of geometric correction or registration of an image is to first obtain apolynomial regression function by control points and then determine the pixel’s locations andbrightness values in the image. A usual method to solve such regression models is ordinaryleast squares (LS) estimation. However, LS estimators are bised and cannot accuratelypropagate and estimate the errors in the image when there are errors in the control points. Theerrors-in-variables (EIV) model is introduced to deal with this problem. It can correct errorscontained in the ground reference control points (RCP) and track error propagation from theRCP to the corrected image. The objective of this paper is to introduce some new EIV-basedestimators into geometric corrections. Our main work is as follows:1. We review the sources and existing models of geometric correction. According to theprocess of geometric correction, this paper summarizes the main uncertainty factors affectingthe accuracy of geometric correction, such as RCP data collections, the number and the distribution of RCP, correction models and parameter estimation, spatial transformation, andresampling methods. The uncertainty expression and propagation problems in remote sensingimage processing are then analyzed;2. Based on the polynomial regression model and the EIV model, the geometriccorrections are introduced. Some estimation principles are discussed in detail, which are theordinary least squares (LS) estimator, consistent adjusted least squares (CALS) estimator andrelaxed consistent adjusted least squares (RCALS) estimator. To demonstrate the performanceof these estimators, simulation experiments are carried out by artificially generating distortedimage and controlling the levels of errors. Furthermore, the differences among LS, CALS andRCALS methods are compared by correcting a SPOT5remote sensing image. Theeffectiveness of the EIV model is shown in correcting and propagating errors in the RCP;3. On the basis of EIV model, we introduce new geometric correction approaches basedon total least squares (TLS) and scaled total least squares (STLS) estimators. Usingsimulation images and a real Beijing-1remotely sensed imagery, we analyze theperformances of the proposed methods. In comparison with the correction accuracy of LS andTLS methods, we find that STLS method can provide better capability to estimate the modelparameters and correct the errors in RCP. In addition, we have yet comprehensively anddeeply analyzed the impact of the number and accuracy of RCP on the accuracy of geometriccorrections; 4. As for the different accuracy of control points, we introduce a robust geometriccorrection method based on weighted total least squares (WTLS) estimator. Its performacehas been validated by a simulation experiment. In fact, it has been shown that theWTLS-based geometric correction method can improve the accurcy and robustness of thecorrection and is more suitable for actual applications.