A-contrario structural inference for space object detection and tracking

Abstract

Monitoring space assets for safety and sustainability compliance requires active measurement of positional information, commonly done via optical remote sensing in which image-frame coordinates are extracted from a high-noise environment of stellar, atmospheric, and hardware features. This task has been traditionally approached by using a-priori models to differentiate potential Anthropogenic Space Objects from the background noise. Source extraction and track-before-detect methods rely on absolute pixel intensity thresholding and require substantial processing to remove noise (stars, hot pixels, etc.), while machine learning shows promise in reducing processing and improving low-visibility performance but requires context-specific labeled training data. We introduce a new approach based on a-contrario detection, arguing that any space object must be unattributable to noise using a sequence of low-fidelity hypotheses. Through this approach, we aim to relax the dependency on a-priori assumptions of data content and improve performance where high-quality data is sparse, poorly labeled, or challenging to characterize, e.g., in satellite-based applications, and provide detection confidence measures from individual data content to enhance risk evaluation for orbital populations. We present an initial qualitative proof-of-concept for our a-contrario approach using data collected by the ASTRIANet telescope network, showing potentially strong performance on Medium Earth Orbit observations for both rate and sidereal-tracking telescope modes. We also discuss how our approach handles epistemic uncertainties, i.e., a lack of a-priori model information, with implications to Type I and Type II error sources and potential mitigation steps when considering Low Earth Orbit observations with higher tracking noise.