You are on page 1of 2

Aircraft Detection for Remote Sensing Images:

A comparative evaluation of object detection using


Ultralytics-YOLO
1st Huzair Saputra 2nd M. Haekal Alfanshury 3rd Deny Joefakri Iwa Supriatna
Master of Artificial Intelligence, Master of Artificial Intelligence, Master of Artificial Intelligence,
Department of Informatics, Universitas Department of Informatics, Universitas Department of Informatics, Universitas
Syiah Kuala, Aceh, Indonesia Syiah Kuala, Aceh, Indonesia Syiah Kuala, Aceh, Indonesia
huzairsaputra@gmail.com haekalalfanshury@gmail.com denyjoefakri@gmail.com

4th Farrasa Rani Faisyal 5th Kahlil Muchtar*


Master of Artificial Intelligence, Department of Electrical and Computer
Department of Informatics, Universitas Engineering, Universitas Syiah Kuala
Syiah Kuala, Aceh, Indonesia Telematics Research Center (TRC)
farasa.lldikti13@gmail.com Universitas Syiah Kuala, Aceh,
Indonesia
kahlil@usk.ac.id (corr.author)

Abstract— Aircraft detection in remote sensing images This study aims to conduct a comparative evaluation of
plays a crucial role in various applications, such as aerial aircraft detection using Ultralytics-YOLO in remote sensing
surveillance, border security, and environmental images. The primary objective is to assess the performance
monitoring. In recent years, deep learning-based object and efficacy of Ultralytics-YOLO in detecting aircraft, as
detection methods have shown remarkable performance well as to compare it with other state-of-the-art object
in identifying and localizing objects in complex visual detection methods. To achieve this, a comprehensive dataset
scenes. Among these methods, Ultralytics-YOLO (You of remote sensing images encompassing diverse aerial scenes
Only Look Once) has gained significant attention for its will be utilized. Additionally, the execution time required for
real-time performance and high accuracy. This paper aircraft detection will be measured to assess the real-time
presents a comprehensive comparison of all variants of capabilities of Ultralytics-YOLO [2].
YOLOv5 and YOLOv8, two popular object detection In summary, this study aims to evaluate and compare
frameworks. The objective is to evaluate their the performance of Ultralytics-YOLO in aircraft detection,
performance and identify the strengths and weaknesses of providing valuable insights into its capabilities and
each variant. A diverse range of datasets, namely RSOD- limitations. The subsequent sections will present the related
aircraft, is used for evaluation. The experimental setup work, proposed approach, results, and conclusion.
involves configuring and training each variant, followed
II. RELATED WORK
by performance evaluation using standard metrics such
as precision, recall, and mean Average Precision (mAP). Previous research by Berkay Yaban, et al. in aircraft
The results highlight the performance differences and detection in very high-resolution satellite images has gained
trade-offs between the variants, aiding researchers and significant attention due to its potential applications in aerial
practitioners in selecting the most suitable model for surveillance, military reconnaissance, and aviation safety.
object detection tasks. The findings contribute to the Deep learning methods, such as the YOLO (You Only Look
advancement of object detection research and facilitate Once) architecture, have demonstrated promising
informed decision-making in various domains. performance in object detection tasks, including aircraft
detection. In this section, we review related works that have
Keywords— Aircraft detection, remote sensing images, employed YOLO-based deep-learning methods for aircraft
object detection, Ultralytics-YOLO detection in very high-resolution satellite images [3].

I. INTRODUCTION
III. PROPOSED APPROACH
In recent years, deep learning-based object detection
methods have shown remarkable advancements in addressing This study suggests an Ultralytics-YOLO object
these challenges. One such approach that has gained attention detection method for use on aircraft. A total of 446 photos
is Ultralytics-YOLO (You Only Look Once), a state-of-the- with 4993 aircraft objects make up the dataset, called RSOD-
art object detection algorithm known for its real-time Dataset [4]. Examples of these images can be seen in Fig. 1.
performance and high accuracy. Ultralytics-YOLO has been
successfully applied in various domains, including computer
vision and robotics. However, its performance and
effectiveness in aircraft detection within remote sensing
images have not been extensively evaluated and compared to
other object detection algorithms [1].
performance in real-world scenarios and identify areas for
improvement. By applying the model's prediction function to
detect all aircraft in an image, we could evaluate its accuracy
and efficiency in identifying these specific objects.

Fig.1. Example of aircraft image


In this paper, we propose a comprehensive evaluation to
compare all variants of YOLOv5 and YOLOv8, including
Yolov8n, Yolov5n, Yolov8s, Yolov5s, Yolov8m, Yolov5m,
Yolov8i, Yolov5i, Yolov8x, and Yolov5x, respectively. The
objective is to assess their performance and identify the
strengths and weaknesses of each variant in the context of
object detection. To conduct this comparison, we carefully
selected the RSOD aircraft dataset as it provides a
challenging and diverse set of images containing various
aircraft instances. The dataset encompasses different aircraft Fig.2. Comparison of model performance YOLOv5 and
types, poses, and environmental conditions, ensuring a YOLOv8 model
comprehensive evaluation of the models. The object prediction result can be seen in Fig. 3.
For each YOLOv5 and YOLOv8 variant, we configured
and trained the models using the appropriate training
strategies specific to each variant. This involved setting up
the network architectures, adjusting hyperparameters, and
initializing the models with suitable weights. To evaluate the
performance of each variant, we utilized standard evaluation
metrics for object detection, including precision, recall, mean
Average Precision (mAP), and F1-score. Additionally, we
considered class-specific metrics to evaluate their
performance on different aircraft categories present in the
dataset. Fig.3. The detection results of YOLOv5x model
Our comparative analysis revealed valuable insights
into the strengths and weaknesses of each YOLOv5 and V. CONCLUSION
YOLOv8 variant. We thoroughly examined their The results emphasize the positive impact of using the
performance in terms of accuracy, computational efficiency, YOLOv8 model for object detection tasks, showcasing
robustness to occlusion, and ability to handle variations in significant improvements over previous versions. This
lighting conditions and object scale. The analysis provided a knowledge can guide future research and development efforts
comprehensive understanding of the performance differences to enhance object detection capabilities, leading to more
and trade-offs between the variants. Based on the results and efficient and accurate systems in various applications.
analysis, this study offers guidance for selecting the most
suitable YOLOv5 or YOLOv8 variant for specific object ACKNOWLEDGMENT
detection applications, particularly in the domain of aircraft We thank the Master of Artificial Intelligence Study
recognition. Program and Erasmus+ DS&AI Program, for supporting this
research.
IV. EXPERIMENTAL RESULTS
REFERENCES
Overall, the results show a significant improvement in
object detection using the YOLOv8 variant model trained [1] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “YOLOv4: Optimal
Speed and Accuracy of Object Detection.” 2020.
with images, but the highest mean Average Precision (mAP) [2] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You Only Look
score among the YOLO object detection models listed is Once: Unified, Real-Time Object Detection,” in 2016 IEEE
achieved by YOLOv5x, with an mAP of 0.742. The Conference on Computer Vision and Pattern Recognition (CVPR),
performance comparison between the model trained with 2016, pp. 779–788. doi: 10.1109/CVPR.2016.91.
[3] B. Yaban, U. Alganci, and E. Sertel, “Aircraft Detection in Very High
original images can be seen in Fig. 2. The mAP scores for Resolution Satellite Images using YOLO-based Deep Learning
different versions of YOLO object detection models. Methods,” in 4th Intercontinental Geoinformation Days (IGD), Tabriz,
We conducted an analysis using the YOLO prediction Iran, May 2022, pp. 270–273.
[4] Y. Long, Y. Gong, Z. Xiao, and Q. Liu, “Accurate Object Localization
model to gain further insights into the model's performance
in Remote Sensing Images Based on Convolutional Neural Networks,”
and understand the areas of focus during the object detection IEEE Transactions on Geoscience and Remote Sensing, vol. 55, no. 5,
process. This analysis provided valuable information on the pp. 2486–2498, 2017, doi: 10.1109/TGRS.2016.2645610.
model's strengths and limitations, allowing us to assess its

You might also like