Automatic Registration of Remotely Sensed Images by Using SURF Features and RANSAC Algorithm

Intensive work has been carried out for optimization of automatic registration using remotely sensed data and photogrammetric techniques because of very large and various datasets availability. Automatic registration is used in many remote sensing mapping applications such as environmental monitoring, change detection, classification, image fusion, etc. In this study, a feature based approach was proposed for automatic registration which can be used for automatic registration of multispectral images acquired in different periods. This technique suggests an optimization of multiband spectral data generated by PCA (Principal Component Analysis) transformation. The multispectral image data was first evaluated using PCA then the SURF (Speeded up Robust Feature) algorithm was applied on the optimized first band of the processed image to detect interest points. In order to decide on matching points used SSD (Sum of Square Distances) values are calculated using interest points data with 64 dimensional feature vectors. As a step forward weak points were eliminated by applying RANSAC (Random Sample Consensus) method and the remaining point data were used for determining homography which is necessary for projective transformation. In the last step, georeferenced images that were geometrically transformed using homography matrix were saved after resampling process. In order to test the proposed approach multispectral aerial images from 2003, 2008 and 2015 were used. The orthophoto image of 2015 was used as reference data. As a result spatial accuracies were found with RMSE values as ± 0.61m and ± 0.53m for the years 2003 and 2008 respectively