The rapid proliferation of artificial satellites poses a growing challenge to astronomical observations, calling for robust methods to flag and mitigate their impact on scientific data quality. Modern astronomical surveys, including space-based missions such as CHEOPS, generate vast data volumes, where the manual identification of these contaminants is unfeasible. Artificial intelligence (AI) has emerged as an essential tool for efficiently processing these large datasets, enabling the automated flagging of transient features to preserve the scientific value of the data. We developed and validated a computationally efficient AI algorithm, based on the MobileNetV2 architecture, to detect satellite trails in CHEOPS images. We benchmarked this novel method against traditional linear feature detection algorithms to assess trade-offs in terms of sensitivity and speed. We trained a binary classifier using an iteratively enhanced dataset, incorporating "hard-negative" examples (e.g., cosmic rays, stray light, Earth limb proximity) to minimize the false-positive rate. The final model was applied to the entire CHEOPS archive of 1.8 million images (up to June 2025). The detections were cross-matched with the Space-Track database to identify objects, enabling a detailed analysis of their physical parameters and magnitude evolution over time. The AI model achieved 99.2% accuracy on the test set and identified 12,223 satellite trails in the archive (0.68% of all images), more than double the yield of non-AI methods, demonstrating superior sensitivity to faint trails. The post-processing identification matched these trails to 5,565 distinct objects. While our photometric analysis from 2020 to 2025 shows a constant average standard magnitude (13.4± 1.7) for the aggregate detection set, an analysis against launch dates reveals a trend of newer objects appearing brighter. AI-based methods provide a powerful and sensitive tool for detecting satellite trails in space-based observatories. However, they do require careful training to generalize against complex image artifacts.
García-Martín et al. (Wed,) studied this question.