PERFORMANCE EVALUATION OF LEARNING-BASED METHODS FOR MULTISPECTRAL SATELLITE IMAGE MATCHING

Abstract. Multispectral image registration is one of the most critical requirements to achieve reliable remote sensing goals such as change detection, image fusion, etc., due to providing complementary knowledge of the scene. On the one hand, this issue has always been a hot topic of research according to significant appearance differences, including geometric and nonlinear radiometric distortions. On the other hand, developing deep learning methods promises precise results in image processing and, in particular, image registration. It is no longer limited to low-level information structures, such as intensity and gradients. However, it is possible to provide more reliable results by extracting various high-level features and removing feature engineering. Therefore, we need extensive experiments in multispectral image registration to determine an efficient and robust method. To this end, this paper evaluates six well-known recently proposed learning-based feature descriptors, including LOFTR, TFeat, HardNet8, HardNet, SosNet, and HyNet, against geometric distortions within real multispectral images. Evaluations demonstrate the general superiority of the HardNet8 descriptor due to extracting high-level features within eight convolution layers.