<![CDATA[Automated military aircraft detection from aerial imagery is a key enabler for modern Intelligence, Surveillance, and Reconnaissance (ISR) operations, particularly for nations with vast airspace such as Indonesia. While the You Only Look Once (YOLO) family has become a de facto standard for such tasks, the literature lacks a controlled, head-to-head comparison of its recent generations on a standard military aircraft benchmark, which leaves architecture selection for ISR systems largely guided by version recency rather than empirical evidence. This study addresses that gap by benchmarking three YOLO generations YOLOv5s, YOLOv8s, and YOLOv11s on the MAR20 dataset (3,842 images, 20 classes, 22,341 instances) under identical training conditions, and by analyzing the results along three axes: aggregate accuracy, per-class behavior, and accuracy–parameter trade-off. Counter to the assumption that newer is necessarily better, the oldest architecture in the comparison, YOLOv5s, achieves the highest Precision (0.985) and F1-Score (0.978) with the smallest parameter count (7.2M), while YOLOv11s, the most recent iteration, leads only on Recall (0.972) and YOLOv8s offers the most balanced profile; all three exceed mAP@0.5 of 0.988. Per-class analysis further shows that detection difficulty is driven by inter-class visual similarity rather than class imbalance, with visually similar fighter aircraft (F-22, F-16, F-15) forming a distinct hard-class cluster (mean AP 0.705–0.788) despite adequate training samples. Based on these findings we propose an operationally-mapped model-selection guide YOLOv5s as a strong candidate for edge-constrained ISR platforms, YOLOv8s for balanced ground-station workloads, and YOLOv11s for recall-critical wide-area surveillance while explicitly noting that the recommendations rest on accuracy–parameter evidence rather than on-device latency measurements, and that the single-seed, small-variant setting bounds the generality of the ranking. Validation through repeated runs with statistical testing, profiling on representative edge hardware, and evaluation on operational imagery remain important directions for future work.]]>
