Enhancing classification accuracy via registration of discrete return LiDAR and aerial imagery using the Levenberg-Marquardt nonlinear optimization method

Description and quantification of a landscape or scene can be achieved by assessing its spectral and structural properties. Fusion of spectral information from aerial imagery and 3-D structural information from LiDAR point clouds allows us to integrate these two complementary characteristics. However, in any fusion method, alignment of data sets is crucial. We registered aerial color (RGB) imagery with LiDAR data by computing a homography matrix(H), using the Levenberg-Marquardt nonlinear optimization method. The root mean square error (RMSE) of registration was less than 0.5 m. The overall classification accuracy of our fusion based object extraction algorithm was also increased from 85% to 90%, when applied to a pre and post registered data set, respectively. In this paper, two different regions were selected to demonstrate the registration method and improved classification results.