ACCESS
Research Article
Volume 2, Article ID: 2025.0027
Mohamed Zarboubi
mohamed.zarboubi@edu.uiz.ac.ma
Abdelaaziz Bellout
abdelaaziz.bellout@edu.uiz.ac.ma
Samira Chabaa
s.chabaa@uiz.ac.ma
Azzedine Dliou
a.dliou@uiz.ac.ma
1 Systems Engineering and Decision Support Laboratory (LISAD), National School of Applied Sciences, IBN ZOHR University, Agadir, 80000, Morocco
2 Laboratory of Systems Engineering and Information Technology (LISTI), National School of Applied Sciences, Ibn Zohr University, PO Box 1136, Agadir, Morocco
3 Instrumentation, Signals and Physical Systems (I2SP) Team, Faculty of Sciences Semlalia, Cadi Ayyad University, Marrakech, Morocco
4 Innovation in Mathematics and Intelligent Systems (IMIS) Laboratory, Faculty of Applied Sciences, Ibn Zohr University, Agadir, Morocco.
* Author to whom correspondence should be addressed
Received: 02 May 2025 Accepted: 03 Dec 2025 Available Online: 11 Dec 2025
Effective pest management remains a persistent challenge in precision agriculture, particularly due to the difficulty of accurately detecting small insects in cluttered trap environments under varying lighting and background conditions. Traditional methods often require labor-intensive inspections, while existing deep learning-based object detectors, such as standard YOLO models, face trade-offs between detection accuracy and computational feasibility on edge devices. This study introduces an improved version of the YOLOv5m model tailored for deployment on a Raspberry Pi-based smart insect trap, targeting codling moth (Cydia pomonella) detection. The proposed architecture incorporates a Convolutional Block Attention Module (CBAM) to enhance feature representation and reduce background interference, along with strategic filter reduction to lower computational complexity. As a result, the model achieves a maximum confidence level of 95% and an average of 91.36%, with reduced parameter count and a FLOPs value of 26.88 billion. Integration with the Firebase IoT platform enables real-time monitoring and remote data management. Comparative analysis with YOLOv5–YOLOv12 variants demonstrates that the improved YOLOv5m offers the best balance between accuracy and efficiency for low-power deployment. These findings highlight the potential of combining lightweight deep learning and IoT infrastructure to create scalable, energy-efficient, and sustainable pest detection systems for real-world agricultural applications.
Disclaimer : This is not the final version of the article. Changes may occur when the manuscript is published in its final format.