Artificial intelligence at the service of space astrometry: a new way to explore the solar system

Advancements in artificial intelligence (AI) have opened new horizons for space exploration, particularly in the domain of astrometry. This research investigates the integration of AI techniques, specifically deep neural networks, with space astrometry using the Cassini-Huygens images database. The primary objective is to train a neural network for the detection and classification of astronomical sources, in particular stars, satellites and cosmic rays, in order to process them for a better understanding of our solar system, performing an analysis on the spatial and temporal variation of cosmic rays and moreover to a wider investigation on the characteristics of the Saturn system.

The methodology employed involves a multi-step process. Firstly, known stars and satellites' positions are located in the images using ARAGO (IMCCE), a software package designed for the astrometric measurement of natural satellite positions in images taken using the ISS of the Cassini spacecraft, resulting in more than 13,000 images (8bit, 1024x1024, exposure duration <= 1 s) correctly calibrated. Consequently a personalised detection system, using classical image processing techniques such as mathematical morphology, is applied to identify all the bright sources within the images, subsequently forming a labeled database for every image including source positions, bounding boxes and corresponding classes—divided in stars, satellites, cosmic rays. 

The database is used to train a YOLOv5 [1] architecture, customised for small object detection, enabling the accurate identification and classification of sources within Cassini images. Due to the different characteristics of the images in the dataset, having a robust detection algorithm is challenging, considering that there is not a ground truth of the signal. It is therefore very hard to detect every source on the images. In order to minimize this issue, every image is divided in four sub-images with a small overlap between them, leading to two main benefits: firstly there is a more constant background on the sub-images allowing an easier background estimation that leads to an easier detection of sources and secondly smaller images are preferable for YOLO training, being optimised for images of size ~ 640x640.  

After those targeted modification the overall precision of the network is ~ 90%. 

The attained outcome is satisfactory for the moment in the context of characterizing cosmic rays’ behavior and in particularl their interaction with Saturn's magnetosphere. Preliminary findings suggest a direct relationship between cosmic rays occurrence and distance from Saturn, with a peak observed in the area outside the magnetosphere’s edge, followed by a decline in cosmic ray incidence with increasing distance from the outer edge of the magnetosphere. Significantly heightened cosmic ray levels has been noted within Saturn's closest proximity (3-8 Rs apart), potentially linked to Enceladus’ plasma emissions [2]. Indeed at Saturn, neutral atoms dominate over the plasma population in the inner magnetosphere, and local source/loss process dominate over radial transport out to 8 RS .

In conclusion, the fusion of artificial intelligence and space astrometry, as demonstrated in this study, introduces a promising paradigm for the exploration of the universe and in particular our solar system. 

[1] ‘You Only Look Once: Unified, Real-Time Object Detection’:  https://arxiv.org/abs/1506.02640

[2] Bagenal, F. (2011) Flow of mass and energy in the magnetospheres of Jupiter and Saturn