The study is devoted to determining the most efficient YOLO-based architecture for the task of aircraft detection in high-resolution aerial imagery. A comparative analysis was conducted across YOLO models v8 through v11 under three experimental conditions: using pre-trained (raw) models, fine-tuning the models on a domain-specific dataset, and fine-tuning models to a dataset enhanced through a proposed image preprocessing method. The evaluation considered both accuracy and inference performance metrics. The proposed methodology reduced the false negative rate from 19.5% to 3.2% at a confidence threshold of 0.75, underscoring its effectiveness in enhancing target visibility under challenging imaging conditions such as low contrast or background clutter.
Aircraft detection in aerial and satellite imagery has received growing attention in remote sensing and computer vision, mainly due to its strategic significance in civilian and defense-related contexts. Traditional approaches relied on handcrafted features and classical machine learning techniques, such as support vector machines, Haar-like features, or HOG descriptors. However, these methods typically struggle to generalize across varying imaging conditions and object appearances.
With the emergence of deep learning, convolutional neural networks have revolutionized object detection, enabling the automatic extraction of hierarchical features directly from raw image data. Region-based detectors, such as Faster R-CNN, and single-shot detectors, such as SSD and YOLO, have demonstrated superior accuracy and speed across diverse visual tasks. Among these, YOLO models have gained traction due to their unified detection pipeline, which allows for real-time inference without compromising accuracy. Numerous studies have employed different YOLO versions for object detection tasks, demonstrating their effectiveness in terms of speed and accuracy. In aircraft detection, YOLO-based models have been adopted to address challenges such as cluttered backgrounds, occlusion, and visually similar non-target objects.
In the study [
U-YOLO is an enhanced detection architecture [
Techniques and innovations developed for ship detection in satellite imagery can be adapted for
aircraft detection tasks, owing to the similar operational conditions and challenges encountered in
satellite-based object detection, such as small target size, low contrast, complex backgrounds, and
varying illumination. In the study [
Despite the availability of a wide range of advanced technologies and object detection models, aircraft detection in high-resolution satellite imagery continues to present significant challenges, including low contrast between aircraft and background, small object size, complex scene composition, and visually similar non-target objects. While existing YOLO-based methods have demonstrated promising speed and detection accuracy results, they often rely on high-quality image inputs and perform suboptimally under adverse visual conditions.
The aim of this research is to determine the most effective YOLO-based architecture for the task of aircraft detection in high-resolution aerial imagery. To this end, a systematic comparative analysis involves four recent YOLO versions: YOLOv8 through YOLOv11—evaluated across key performance metrics, including detection accuracy, inference speed, and robustness under diverse imaging conditions. Beyond baseline assessment, the study introduces an enhanced preprocessing pipeline integrating histogram equalization and bilateral filtering. This preprocessing strategy aims to improve image contrast and edge preservation, thereby enhancing the overall quality of input data and potentially increasing detection performance across all evaluated models.
To enhance the confidence and reliability of aircraft detection in remote sensing imagery, we propose the methodology shown in Figure 1.
The first stage consists of digital image and dataset acquisition. In this study we used the Military
Aircraft Recognition dataset [
The second stage of the proposed methodology involves applying preprocessing techniques
specifically designed to enhance the visual quality of the input data, thereby improving object
detection performance, particularly in low-contrast or noisy conditions [
( ) (‖ ( ) − ( )‖) (‖ − ‖) , ( ) = = 1 ∈ ∈ (1) (‖ ( ) − ( )‖) (‖ − ‖) , where
is the filtered image, I is the original input image, x are the coordinates of the current pixel to be filtered, Ω is the window centered in x, so ∈ Ω in another pixel, is the range kernel for smoothing differences in intensities, is the spatial (or domain) kernel for smoothing differences in coordinates [14]. proposed preprocessing approach, which combines histogram equalization and bilateral filtering applied to raw aerial imagery. Figure 2a depicts the original image, which typically exhibits common challenges characteristic of remote sensing data, including suboptimal contrast, haze, and subtle illumination variations. These factors can obscure fine details of aircraft and reduce edge distinctness, particularly against complex or shadowed backgrounds. In contrast, Figure 2b shows the image after histogram equalization, demonstrating a marked enhancement in overall contrast. This global contrast adjustment redistributes intensity values across the full dynamic range, rendering previously faint features more discernible. However, while histogram equalization improves contrast, it may also amplify noise or introduce undesirable artifacts. Figure 2c presents the result after bilateral filtering, which effectively smooths homogeneous regions while preserving critical edge and structural information. This edge-preserving smoothing is essential for aircraft detection, as it sharpens aircraft contours without blurring their distinct shapes or causing them to merge with the background. Collectively, the combined preprocessing steps yield a perceptibly clearer and more defined aerial scene, wherein aircraft outlines become more prominent and distinguishable from the surrounding environment. Its enhanced visual clarity constitutes a foundational improvement that may contribute to better detection performance by YOLO models in subsequent processing stages.
The third stage involves training various YOLO models on original and preprocessed datasets. Since its inception, the YOLO object detection framework has undergone substantial architectural evolution, with successive versions introducing innovations to enhance detection accuracy, computational efficiency, and task versatility. YOLOv8, released by Ultralytics in 2023, represents a significant advancement by adopting an anchor-free detection paradigm, simplifying the architecture, and improving performance in detecting small-scale objects. It supports a unified framework for multiple vision tasks, including object detection, instance segmentation, and classification, and extends applicability to object tracking, pose estimation, and oriented bounding boxes [15].
YOLOv9 (2024) further introduces the Programmable Gradient Information mechanism and the Generalized Efficient Layer Aggregation Network, addressing information bottlenecks and improving lightweight model performance without sacrificing inference speed [16]. YOLOv10, also from 2024, implements a non-maximum suppression (NMS)-free training paradigm that enables endto-end deployment and reduces computational costs. Additional architectural enhancements, such as spatial-channel decoupled down sampling and large-kernel convolutions, contribute to increased accuracy and efficiency, while a dual assignment strategy lowers latency for real-time detection [17].
The latest iteration, YOLOv11, continues this trajectory by integrating novel components, including the C3k2 block, Spatial Pyramid Pooling-Fast, and a Convolutional Block with Parallel Spatial Attention. These advances substantially improve feature extraction and object localization, enhancing performance across various vision tasks such as oriented object detection and instancelevel segmentation [18, 19].
The fourth stage evaluates model performance using a comprehensive set of metrics. The metrics computed include Precision, Recall, mAP50, and mAP50-95 to quantify detection accuracy. In addition to these accuracy-oriented measures, efficiency metrics such as Frames Per Second and Latency were assessed to characterize the models’ real-time inference capability and response time, respectively. Confusion matrices were constructed to provide deeper insights into classification performance, enabling detailed visualization of true positives, false positives, false negatives, and true negatives. The evaluation also includes a comparative analysis of models trained on preprocessed versus non-preprocessed datasets, thereby elucidating the impact of the preprocessing pipeline on detection performance, particularly in challenging conditions characterized by low contrast or image noise.
Finally, the optimal YOLO model architecture was selected based on a thorough trade-off analysis considering detection accuracy, inference speed, and robustness to low-contrast imagery. The chosen model represents the best overall compromise between Precision, Recall, computational efficiency, and reliability, thus constituting the most effective approach for accurate and efficient aircraft detection in high-resolution remote sensing imagery.
The experiment aimed to evaluate and compare the performance of four YOLO model versions – YOLOv8, YOLOv9, YOLOv10, and YOLOv11 – on the task of aircraft detection using aerial imagery and compare them in the following states: not fine-tuned, fine-tuned, fine-tuned on the dataset with the proposed method applied. To conduct the experiment, the software was created in the Python programming language, using Ultralytics YOLO library for model training. The experiment was conducted on the following setup: CPU AMD Ryzen 5600X, GPU: MSI RTX 3080 Gaming X Trio, RAM: 32GB.
The mean Average Precision (mAP) metric is employed for comprehensive assessment as YOLObased models evaluate performance based on object classification and localization accuracy. The mAP metric integrates Precision and Recall over a range of confidence thresholds and is computed at varying Intersections over Union (IoU) thresholds. These thresholds define the minimum overlap required between the predicted and ground truth bounding boxes for a detection to be considered a True Positive. Varying the IoU threshold affects the classification of detections as True Positives or False Positives, thereby influencing the resulting Precision and Recall values. Precision is defined as the ratio of True Positive detections to the total number of positive detections made by the model at a given IoU threshold and is computed according to Equation 2: where TP is true positive, FP is false positive.
As defined by Equation 3, Recall represents the proportion of True Positive detections identified by the model relative to the total number of ground truth instances. It quantifies the model’s ability to detect all relevant objects at a given threshold correctly: where TP is true positive, FN is false negative.
The general formulation for calculating mAP is provided in Equation 4: where N is the total number of object classes, and APi denotes the Average Precision for the i-th class.
This metric reflects the meaning of the precision-recall integrals computed for each class, providing a single detection accuracy measure across multiple categories.
Specifying the IoU thresholds used in mAP calculations with numeric suffixes is customary. For instance, mAP50-95 indicates that the mAP was computed by averaging AP values over IoU thresholds ranging from 0.50 to 0.95 in increments of 0.05. In contrast, mAP50 refers to the AP value calculated at a single IoU threshold of 0.5.
The Average Precision (AP) for a given object class is computed by first ranking all predicted bounding boxes in descending order according to their confidence scores. Following this, Precision and Recall values are calculated at each detection threshold. The AP is then obtained by computing the area under the precision-recall (P–R) curve, which is estimated through numerical integration over all Recall levels. This process is formally expressed in Equation 5:
, , , = where p(r) represents the Precision as a function of Recall.
To assess the inference performance of each model, two key metrics were employed: Frames Per Second (FPS) and Latency. FPS measures the number of image frames that a model can process within one second, providing an indicator of real-time processing capability. In contrast, Latency quantifies the time required by the model to process a single image, typically expressed in milliseconds. Together, these metrics offer a comprehensive evaluation of the computational efficiency and responsiveness of each YOLO variant during inference.
In the fields of machine learning and computer vision, the confusion matrix is a fundamental tool for evaluating the performance of classification models [20]. It provides a comprehensive visualization of classification outcomes by quantifying the number of correct and incorrect predictions across all categories. Specifically, it reports the count of true positives, which are correctly identified positive instances, and true negatives, which are correctly identified negative instances. False positives correspond to negative samples that are incorrectly classified as positive, while false negatives denote positive samples that are mistakenly classified as negative. By capturing these four key components, the confusion matrix facilitates a detailed assessment of a model's predictive accuracy and error distribution.
The performance of the non-fine-tuned YOLO models on the aircraft detection task, shown in Table 1, demonstrates strong baseline capabilities. Specifically, the models achieved high Precision values ranging from 0.90 to 0.96 and Recall values between 0.89 and 0.94. The mAP50 scores exceeded 0.94 across all models, indicating effective localization of aircraft in most cases. The models exhibited competitive performance in inference efficiency, with frame rates between 75 and 87 FPS and Latency values ranging from 5.67 to 7.92 milliseconds per image, depending on the model.
However, the mAP50-95 metric provides a more comprehensive evaluation of detection accuracy across varying IoU thresholds — remained comparatively low, ranging from 0.65 to 0.69. It suggests that while the models identify object presence and location well, their ability to generalize across stricter localization criteria without fine-tuning is limited.
According to Table 2, the experimental results indicate minimal variation in detection accuracy across the evaluated fine-tuned YOLO models when applied to the aircraft detection task. Specifically, the mAP50 remains nearly identical across all versions, suggesting comparable performance in terms of basic object localization.
In terms of inference efficiency, YOLOv11m outperformed all other models. It achieved the lowest Latency at 5.66 milliseconds per image and the highest throughput with 87.82 FPS. The next bestperforming model in this regard was YOLOv8m, with a Latency of 6.34 milliseconds per image and 86.31 FPS. Visual comparison charts for performance metrics are shown in Figure . 6 , 0 8 6 7 , 6 5 F P S
An additional aspect considered in the evaluation was the number of trainable parameters directly impacting model size and deployment feasibility. The models exhibited noticeable differences: YOLOv8m contains 25.8 million parameters, YOLOv9m and YOLOv11m each have 20.1 million, while YOLOv10m is the most lightweight, with only 16.5 million parameters. YOLOv10m maintains competitive performance despite its lower parameter count, suggesting an efficient architectural design. Confusion matrices for the evaluated fine-tuned YOLO models are presented in Figure 4, computed using a confidence threshold of 0.25 and an IoU threshold of 0.45. These results demonstrate that YOLOv11m offers improved accuracy under stricter evaluation criteria and superior computational efficiency while maintaining average trainable parameter numbers across models.
A visual comparison of detection outputs from the evaluated YOLO models is presented in Figure 5, using a test image characterized by moderate contrast between aircraft and background. Notably, the plane in the scene exhibits color tones similar to the surrounding environment, which poses a challenge for object-background discrimination. The image also contains multiple buildings of varying shapes and sizes, potentially leading to false positive detections due to their structural similarity to aircraft.
All models demonstrate high confidence scores for correctly detected aircraft, particularly larger objects. However, performance degrades for smaller aircraft, with reduced confidence levels and, in some cases, missed detections. Specifically, the confidence scores for a small aircraft in the scene were 0.78 for YOLOv8, 0.68 for YOLOv9, and 0.47 for YOLOv10. YOLOv11 failed to detect the small aircraft, indicating a limitation in its sensitivity to small-scale targets under low-contrast conditions.
The models demonstrated moderate detection performance in another test image, which features low contrast between the aircraft and the background and contains six aircraft (Figure 6). All models successfully detected at least four aircraft, though with varying levels of confidence and accuracy.
YOLOv8 detected all six aircraft. However, two detections were associated with low confidence scores of 0.26 and 0.67, respectively. YOLOv9 correctly identified four aircraft, one with a confidence score of 0.63. YOLOv10 achieved the most consistent performance in this scenario, detecting all six aircraft with confidence scores exceeding 0.70. YOLOv11 detected three aircraft with high confidence values (greater than 0.84), identified a fourth aircraft with a lower confidence of 0.44, failed to detect one aircraft, and produced one false positive detection with a confidence score of 0.42. a)
b)
Table 3 presents the performance metrics of the fine-tuned YOLO models evaluated on the dataset preprocessed using the proposed method. The results indicate that Precision, Recall, and mAP50 values remain comparable to those obtained from the same models fine-tuned on the original, nonpreprocessed dataset (as shown in Table 2). Similarly, the inference speed metrics, namely FPS and Latency, exhibit negligible differences, confirming that the preprocessing pipeline does not introduce computational overhead.
Model Parameters Precision Recall mAP50 mAP50-95 FPS Latency (ms/img) YOLOv8m 25.8M 0.992688 0.985180 0.994364 0.782252 86.45 6.30 YOLOv9m 20.1M 0.993128 0.988755 0.994658 0.785548 75.79 7.94 YOLOv10m 16.5M 0.963750 0.944347 0.984384 0.689463 86.24 6.96 YOLOv11m 20.1M 0.993315 0.985812 0.994526 0.786554 88.12 6.21
Although a slight decrease is observed in the mAP50-95 metric, indicating a modest reduction in bounding box localization accuracy under stricter IoU thresholds, this is offset by an improvement in the confidence levels of the detected objects.
As shown in Figure 7, the proposed technology significantly reduces the number of false negative detection results on higher confidence levels. From 2364 (19.5%) to 449 (3.2%).
All evaluated YOLO models demonstrated consistently high Precision, with the lowest recorded value of 0.964 for YOLOv10 and the highest of 0.993 for YOLOv11. Similarly, Recall values remained robust across models, ranging from 0.944 (YOLOv10) to 0.989 (YOLOv9). The mAP50 exhibited a comparable trend, with values spanning from 0.984 (YOLOv10) to 0.995 (YOLOv9). For the stricter mAP50-95 metric, YOLOv10 again yielded the lowest score at 0.689, while YOLOv11 achieved the highest at 0.787.
Regarding performance, YOLOv9 was identified as the slowest model, with a frame rate of 75.79 FPS and Latency of 7.94 ms/image. YOLOv8, YOLOv10, and YOLOv11 demonstrated comparable performance results, with YOLOv11 emerging as the most efficient, delivering 88.12 FPS and 6.21 ms/image latency.
In conclusion, based on accuracy and performance metrics, YOLOv11 is the most effective architecture for aircraft detection in high-resolution remote sensing imagery. However, YOLOv10 and YOLOv8 also yielded competitive results, and due to their widespread adoption and the availability of extensive modifications and enhancements made by other researchers, these models may provide practical benefits in scenarios where adaptability and extensibility are prioritized.
This study conducted a systematic comparative evaluation of four recent YOLO architectures (YOLOv8 through YOLOv11) to identify the most effective solution for aircraft detection in highresolution satellite imagery. To enhance detection performance under challenging visual conditions, the proposed methodology incorporated a preprocessing stage combining histogram equalization and bilateral filtering. This approach was specifically designed to improve global contrast and suppress background noise while preserving critical edge details, thereby enhancing the quality of the input data provided to the detection models.
This study conducted a systematic comparative evaluation of four recent YOLO architectures (YOLOv8 through YOLOv11) to identify the most effective solution for aircraft detection in highresolution satellite imagery. To enhance detection performance under challenging visual conditions, the proposed methodology incorporated a preprocessing stage combining histogram equalization and bilateral filtering. This approach was specifically designed to improve global contrast and suppress background noise while preserving critical edge details, thereby enhancing the quality of the input data provided to the detection models. The effectiveness of this preprocessing strategy was validated through a comparative analysis of model performance on original versus preprocessed datasets. Notably, it resulted in a substantial reduction in the false negative rate—from 19.5% to 3.2% at a confidence threshold of 0.75—demonstrating its significant contribution to improved object visibility in low-contrast or cluttered scenes. Among the evaluated models, YOLOv11 demonstrated the best overall performance. It achieved the highest Precision (0.993), the second-highest Recall (0.9858), and the best mAP50-95 (0.7866) score, reflecting strong performance in both localization and classification tasks. Furthermore, with a frame rate of 88.12 FPS and latency of 6.21 ms/image, YOLOv11 meets the requirements for real-time inference, making it highly suitable for operational use. The results confirm that integrating a preprocessing stage with the YOLOv11 architecture provides a robust and efficient framework for accurate aircraft detection in high-resolution aerial imagery, particularly under suboptimal imaging conditions.
[13] G. Papari, N. Idowu, T. Varslot. “Fast bilateral filtering for denoising large 3D images”. IEEE
Transactions on Image Processing, vol. 26, no. 1, 2017, pp. 251–261. [14] R. Gavaskar, K. Chaudhury. “Fast Adaptive Bilateral Filtering”. IEEE Transactions on Image
Processing, vol. 28, no. 2, pp. 779-790, Feb. 2019, doi: 10.1109/TIP.2018.2871597. [15] Y. Chen, X. Yuan, J. Wang, R. Wu, X. Li, Q. Hou. “YOLO-MS: Rethinking Multi-Scale Representation Learning for Real-time Object Detection”. IEEE Transactions on Pattern Analysis and Machine Intelligence 47, pp. 4240–4252, 2023. doi: 10.48550/ARXIV.2308.05480. [16] C.-Y. Wang, I.-H. Yeh, H.-Y. M. Liao, YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information, in: Computer Vision – ECCV 2024, Springer Cham, 2024, pp.1–12. doi: 10.1007/978-3-031-72751-1_1. [17] Y. Li, W. Leong, H. Zhang, YOLOv10-Based Real-Time Pedestrian Detection for Autonomous Vehicles, in: 2024 IEEE 8th International Conference on Signal and Image Processing Applications (ICSIPA), Kuala Lumpur, Malaysia, 2024, pp. 1–6. doi: 10.1109/ICSIPA62061.2024.10686546. [18] R. Khanam, M. Hussain, YOLOv11: An Overview of the Key Architectural Enhancements, 2024.
doi: 10.48550/arXiv.2410.17725. [19] L. He, Y. Zhou, L. Liu, W. Cao, J. Ma, Research on object detection and recognition in remote sensing images based on YOLOv11, Scientific Reports 15, 2025. doi: 10.1038/s41598-025-96314x. [20] V. Hnatushenko, D. Mozgovoy, V. Vasyliev, Accuracy evaluation of automated object recognition using multispectral aerial images and neural network, in: Tenth International Conference on Digital Image Processing (ICDIP 2018), Shanghai, China, SPIE, 2018, p. 72. doi: 10.1117/12.2502905.