Enhancing Space Debris Detection and Classification Using Optimized YOLO Variants

With the emergence of satellite-related services, the space surrounding the Earth is being filled with futile objects, such as broken satellite pieces or natural debris. These space debris enhance the likelihood of collision with operational satellites and add to the hazards associated with future endeavors. Thus, implementing deep learning models to localize and identify space objects is crucial for the safety of operating satellites and space stations. In this study, 20 different model variants of YOLOv8, YOLO-NAS, YOLOv9, RT-DETR, and YOLOv10 are investigated for localization using the Space Debris Detection Dataset and SPARK Dataset. The Space Debris Detection Dataset is a single class multi-object dataset, whereas the SPARK dataset comprises 11 classes, of which 10 are spacecraft and the remaining class is space debris. After the datasets have been processed to generate labels that adhere to the YOLO label format, the models undergo training and evaluation. Through performance analysis, YOLOv8-M emerged as the most optional model with an impressive 99.8% precision, 99.5% recall, 99.5% mAP@50, and 82.7% mAP@50-95 on the Space Debris Detection Dataset. Subsequently for SPARK Dataset, YOLOv8-M and YOLOv9-C portray remarkable classification capability. The findings of this research highlight that YOLOv8-M and YOLOv9-C excel in space object localization and classification with high accuracy and low inference time, charting a path for future developments in real-time debris detection and enhanced orbital safety.