An Empirical Study of YOLOv10 on Optical vs SAR Imagery for Aircraft Recognition

This study investigates the effectiveness of the YOLOv10 object detection algorithm for aircraft recognition using both optical and Synthetic Aperture Radar (SAR) satellite imagery. Leveraging high-resolution datasets acquired from Google Earth and SAR imaging systems, the model’s performance was evaluated using standard object detection metrics, including precision, recall, mean Average Precision (mAP), F1-score, and inference time. The objective was to assess the robustness and generalizability of YOLOv10 across distinct image modalities that vary significantly in spectral and structural content. Experimental results demonstrate that YOLOv10 achieves superior performance on optical imagery, with high detection accuracy and reliable localization of aircraft in complex airport environments. Quantitatively, the model attained a precision of 0.859, a recall of 0.853, an mAP of 0.813, and an F1-score of 0.856 on optical images. In contrast, SAR imagery posed greater challenges due to speckle noise, lower contrast, and more ambiguous structural features. As a result, the model exhibited decreased performance, achieving a precision of 0.792, recall of $0.787, \mathrm{mAP}$ of 0.766, and an $F 1$-score of 0.789. However, the inference time on SAR imagery was slightly faster, suggesting a computational trade-off. The findings underscore the importance of modality-aware model training in remote sensing object detection tasks. While YOLOv10 demonstrates high effectiveness in optical scenarios, its performance on SAR data indicates a need for targeted preprocessing, domain adaptation, or model fine-tuning to achieve comparable accuracy. These results provide valuable insights for the deployment of real-time aircraft detection systems using heterogeneous satellite imagery.