Fall armyworm (Spodoptera frugiperda) is one of the most serious pests in maize fields. It is often required to early and accurately detect its infestation for timely and effective pest prevention using unmanned aerial vehicle (UAV) imagery. However, reliable detection has been confined to the challenges: 1) The small and subtle feeding marks caused by the larvae, leading difficult to identify at high altitudes. 2) Consistent recognition has been limited to significant variations in object scale at different flight heights. 3) The accurate detection has also been limited to the low contrast between damaged leaf tissue and surrounding healthy foliage, especially under the different lighting and environmental conditions in fields. Collectively, advanced computer vision is necessary to robustly identify early signs of infestation at diverse scales under complex backgrounds. In this study, a robust deep learning model was developed to reliably identify the subtle infestation traces in multi-scale UAV images. A detection architecture, termed coordinated-BiFPN-P2-YOLO (CBP-YOLO), was also proposed using YOLOv8. Real-enhanced super-resolution generative adversarial network (Real-ESRGAN) was applied as a preprocessing step to reduce image degradation from low ground sampling distance. High-fidelity textures of leaf damage were reconstructed from original low-resolution inputs. The backbone of YOLOv8 was enhanced with the Coordinated attention (CA) mechanism. Spatial and channel-wise features were captured to improve the localization and discrimination of minute lesions. Furthermore, the neck component was upgraded with a Bi-directional feature pyramid network (BiFPN) for the highly efficient top-down and bottom-up cross-scale feature fusion. Information loss was minimized for consistent representation during hierarchical propagation at different scales. In addition, a detection head was added to specifically strengthen sensitivity to small targets, particularly at a 160×160 spatial resolution with 64-channel output. The improved model was trained and then evaluated on the custom UAV dataset, which was collected from maize fields naturally infested by fall armyworm under diverse lighting conditions and flight heights. Extensive experiments demonstrated that the CBP-YOLO achieved a peak performance on the imagery with a ground sampling distance (GSD) of 0.38 cm per pixel. Real-ESRGAN significantly alleviated texture blurring and edge ambiguity in low-resolution images, leading to better delineation of feeding scars. Ablation studies were conducted to evaluate the effectiveness of the improved model. There was an outstanding performance on the UAV multi-scale blade dataset. Specifically, there was an average precision (AP@0.5) of 76.5%, which increased by 3.4 percentage points, compared with the baseline model. The robustness and practical applicability of the improved model were obtained in the blades of varying scales during aerial inspection. A comparison showed that the CBP-YOLO outperformed state-of-the-art detectors—including YOLOv9 medium, YOLOv10 medium, YOLOv11 medium, Faster region-based convolutional neural network, and RetinaNet—by margins of 10.1, 7.2, 5.1, 9.3, and 17.9 percentage points in AP@0.5, respectively. Notably, the high precision was also maintained under varying illumination and partial occlusion, indicating strong generalization in agricultural environments. The improved CBP-YOLO framework effectively detected subtle, multi-scale fall armyworm infestation signals during UAV monitoring. Superior accuracy and robustness of the improved model were achieved to synergistically combine super-resolution enhancement, attention-aware feature extraction, fine-grained detection heads, and bidirectional multi-scale fusion. These findings can also provide a practical and scalable solution for early pest outbreak detection, thereby enabling timely intervention to reduce the crop losses in large-scale maize production.