A Machine Learning-Based Semantic Pattern Matching Model for Remote Sensing Data Registration

Remote sensing image registration can benefit from a machine learning method based on the likelihood of predicting semantic spatial position distributions. Semantic segmentation of images has been revolutionized due to the accessibility of high-resolution remote sensing images and the advancement of machine learning techniques. This system captures the semantic distribution location of the matching reference picture, which ML mapped using learning-based algorithms. The affine invariant is utilized to determine the semantic template's barycenter position and the pixel's center, which changes the semantic border alignment problem into a point-to-point matching issue for the machine learning-based semantic pattern matching (ML-SPM) model. The first step examines how various factors such as template radius, training label filling form, or loss function combination affect matching accuracy. In this second step, the matching of sub-images (MSI) images is compared using heatmaps created from the expected similarity between the images' cropped sub-images. Images having radiometric discrepancies are matched with excellent accuracy by the approach. SAR-optical image matching has never been easier, and now even large-scale sceneries can be registered using this approach, which is a significant advance over previous methods. Optical satellite imaging or multi-sensor stereogrammetry can be combined with both forms of data to enhance geolocation.