實驗結果顯示，合成域內準確率不足以作為模型選擇之依據：域內表現最佳之 SPARK2024 於 NASA 真實影像上之遷移表現最弱，以 SPARK2021 訓練之模型則取得最佳 NASA 表現。此差異可由訓練資料與目標域之尺度分布匹配程度解釋（Spearmanρ = −0.905）；且於未使用資料增廣之條件下，多尺度整合之 MS-MERGED 優於所有單一資料集，顯示當目標域分布未知時，多尺度資料整合為較穩健之預設訓練策略。資料增廣之效益亦非固定正向，而受訓練資料尺度分布調制：LD-MERGED 之增廣增益高達 +0.292，LD-MERGED-Cap 受控實驗證實限縮 large object 尺度範圍後，增廣差距即明顯縮小。模型架構方面，RT-DETR 於無增廣條件下展現較佳原生特徵建模能力，YOLOv8n 則於加入增廣後取得最高整體表現，並於 Jetson Orin Nano 邊緣部署中展現最佳之延遲、吞吐量與能源效率。
綜合而言，本研究之貢獻不在於提出單一最佳訓練資料集，而在於建立一套能解 釋並評估 synthetic-to-real 遷移表現之 dataset-aware 與 scale-aware 評估框架，並將模型選擇擴展至準確率、遷移能力與部署效率之三軸權衡，可作為未來光學 SSA 模型設計、跨域驗證與邊緣部署研究之參考。;With the rapid increase of satellites and space debris in low Earth orbit， space object detection has become an important task for Space Situational Awareness (SSA). Optical space imagery typically contains small targets， strong illumination variation， and relies heavily on synthetic training data， making detection performance dependent not only on model architecture but also on dataset distribution， object-scale composition， and deployment constraints. This thesis therefore presents a dataset-aware and scale-aware evaluation framework: YOLOv8n is systematically evaluated across the synthetic SPARK2021， SPARK2022， and SPARK2024 datasets and several scale-grouped merged variants， and compared against RT-DETR variants on SPARK2022; candidate models are then evaluated on a self-annotated NASA real-image test set and deployed on the Jetson Orin Nano.
The results show that in-domain accuracy on synthetic data is an unreliable criterion for model selection: SPARK2024， the strongest in-domain setting， exhibits the weakest transfer to NASA real images， whereas models trained on SPARK2021 achieve the best NASA performance. This difference can be explained by the degree of match between the object-scale distributions of the training data and the target domain (Spearman $\rho$ = −0.905). Notably， without augmentation the multi-scale MS-MERGED setting outperforms all single datasets， indicating that multi-scale data integration is the more robust default strategy when the target-domain distribution is unknown. Augmentation benefit is likewise modulated by the scale distribution of the training data: LD-MERGED exhibits a gain as large as +0.292， and the LD-MERGED-Cap controlled experiment confirms that restricting the large-object scale range substantially narrows this gap. Architecturally， RT-DETR is stronger without augmentation， whereas YOLOv8n achieves the best overall performance once augmentation is applied and delivers the best latency， throughput， and energy efficiency in edge deployment.
Overall， the contribution of this thesis lies not in identifying a single best training dataset， but in establishing an evaluation framework that explains and assesses synthetic-to-real transfer performance， extending model selection to a three-axis trade-off among accuracy， transferability， and deployability. The framework can serve as a reference for future optical SSA model design， cross-domain validation， and edge-deployment research.
