Mid-infrared (MIR) image is highly valued across various fields, such as national defense and environmental monitoring, due to its capability to capture temperatures of objects and surfaces. Image registration that unifies coordinates between images is a fundamental process for utilizing multi-temporal satellite images. The radiation-invariant feature transform (RIFT) algorithm is a feature-based matching method that extracts robust matching points to non-linear radiometric distortions in day and night MIR images. However, the original RIFT method has a limitation in detecting a small number of matching points because the properties of the image are not sufficiently considered. In this study, we propose an optimization method of RIFT for MIR day-night image registration. First, patch size is selected so the RIFT algorithm can stably acquire multiple matching points. After comparing the results of applying RIFT by setting various patch sizes, we select the final patch size that extracts the most matching points. In addition, the RIFT algorithm’s hyperparameters are optimized to suit the characteristics of the MIR image by comparing the number of matching points and RMSE for each combination. Finally, the image registration is conducted using a transformation model based on extracted inlier points by applying random sample consensus (RANSAC), data snooping, and locality preserving matching (LPM), which are outlier removal algorithms. Based on experiments conducted from KOMPSAT-3 MIR day/night satellite images, the LPM algorithm produced the best quantitative evaluation result with an average RMSE and circular error of 90% (CE90) of 0.984 pixels and 2.076 pixels. From the experiments, it was demonstrated that the proposed method can contribute to improving image registration by effectively extracting matching points that reflect the characteristics of KOMPSAT-3A MIR day-night imagery.



Keywords KOMPSAT-3A, Mid-infrared, RIFT, Image registration