24AI Application of SuperPoint/SuperGlue Neural Network Architectures to Improve ABI Solar Channel Calibration Using Lunar Radiance

Measured and modeled lunar irradiances have long been used for vicarious satellite calibration. This project advances a novel approach using lunar radiance, defined as the irradiance per unit solid angle. Three areas of interest (AOIs) were selected from a reference GOES-16 image. A key challenge is projecting these coordinates onto a G-16 subset image when phase angle and libration differ. To address this, the detection and matching of multiple control points across images is essential. Previous methods employed classical computer vision techniques like SIFT. Here, a new algorithm was developed, leveraging robust AI techniques such as Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs). SuperPoint, a self-supervised CNN architecture, was applied on the reference and subset images to extract control points and their descriptors. Unlike traditional methods which require extensive labeled data for training, SuperPoint is trained on synthetic data generated from homographic transformations, allowing it to effectively learn and describe interest points in various conditions. Next, the SuperGlue GNN architecture was applied to match control points. By leveraging the spatial relationships and descriptor similarities from SuperPoint, SuperGlue is able to optimize the matching process. A homography transformation was then calculated from these matches to project AOI coordinates from the reference to the subset image, enabling lunar radiance extraction. Prediction models were constructed using solar/view geometric angles as proxies for lunar radiance.