<p>The detection and classification of aircraft in satellite imagery are critical for applications such as civil aviation monitoring, and surveillance. While recent advances have improved aircraft detection performance, most existing work treats detection and classification as a single end-to-end task and does not explicitly model post-detection misclassification. This study investigates whether features extracted from a YOLO-seg detector can improve system-level performance beyond YOLO alone, particularly for background rejection and reduction of misclassification errors. We propose a framework that explicitly fuses deep visual features with geometric shape descriptors as an independent error-correction stage. Visual embeddings extracted from YOLO, together with confidence scores and polygon-derived geometric metadata, are integrated into an XGBoost classifier to refine object-level predictions without modifying the detector’s backbone. Experiments on the RarePlanes dataset show that the proposed post-classification strategy reduces end-to-end YOLO misclassification errors by up to 51%, with embedding-based features consistently outperforming metadata-only features. Overall, the findings demonstrate that using YOLO-extracted embeddings and metadata provides a practical mechanism for post-detection error correction within a YOLO-based aircraft detection pipeline.</p>