A Hybrid Pipeline for Aircraft Role Classification in Satellite Imagery

The detection and classification of aircraft in satellite imagery are important in various fields such as civil aviation management, security monitoring, and military surveillance. While recent advances in deep learning, particularly object detection models like YOLO, have improved aircraft detection performance, most existing research relies solely on visual features without incorporating structured metadata. By integrating geometric features with deep visual features extracted from a pre-trained YOLO into an XGBoost classifier, this paper demonstrates how multi-modal feature fusion affects classification performance. Through a series of experiments involving YOLO11n-seg and XGBoost models, as well as feature-fusion scenarios with and without metadata, our experimental results show that extracting information from YOLO to feed XGBoost improves overall classification performance, achieving fewer total errors than the YOLO-only model and demonstrating substantial gains in detecting minority classes. Furthermore, applying minority class oversampling techniques improves military class recall but introduces additional noise, highlighting the trade-off between precision and recall.