Weakly Supervised Detection of Satellite Streak Signatures in SST Imagery Using Multiple Instance Learning

Abstract

The number of objects launched into space increases significantly every year. Combined with vast amounts of data produced by imaging instruments each day, automated processing becomes one of the most urgent needs. The manual annotation process, despite being time-consuming and inefficient, is still common. This study proposes the application of multiple instance learning (MIL), coupled with domain-specific preprocessing, as a weakly supervised classification pipeline to support and accelerate the manual annotation process in Space Surveillance and Tracking (SST) imagery analysis. Two feature extraction backbones are evaluated: a pretrained ResNet-18 and a custom convolutional masked autoencoder. Three aggregation methods are compared: max pooling operator as a baseline, attention-based pooling (ABMIL), and CLAM, which involves an additional instance-level classification branch. Heatmap-based interpretability analysis is also conducted to provide insight into the decision-making process of the proposed system. The results demonstrate that the proposed MIL-based pipeline, supported by interpretability methods, effectively addresses key challenges in automated SST image processing, including weak supervision and limited annotated images availability.