Deep Learning-Based Pest Insect Detection for Autonomous Agricultural Systems: Development and Field-Testing of a Custom Convolutional Neural Network Architecture

Selim Sürücü (Çankırı Karatekin University)
Berk Küçük (Çankırı Karatekin University)
Mahammad Karimzade (Çankırı Karatekin University)

Indonesian Journal of Electrical Engineering and Informatics (IJEEI)
Vol 14, No 2: June 2026

Global food security is under increasing threat from pest organisms that target agricultural production. The broad-spectrum pesticides traditionally used to combat these pests cause permanent damage to ecosystems, pose risks to human health, and lead to a critical decline in biodiversity. Precision agriculture technologies, as part of the Agriculture 4.0 revolution, offer innovative solutions to these problems. This study focuses on developing a novel artificial intelligence model capable of detecting and classifying pest insects with high accuracy, to serve as the intelligence for autonomous spraying systems. Within the scope of this research, a comprehensive and diverse dataset of over 12,000 images was meticulously compiled, featuring a total of 11 insect species, including 8 economically significant pests and 3 beneficial species for the ecosystem. To enhance the robustness and diversity of the dataset, advanced Generative Adversarial Networks (GANs) such as StyleGAN2-ADA were used alongside traditional data augmentation techniques to generate synthetic images. On this rich dataset, VGG-16 and ResNet152 V2 models, representing the transfer learning approach, were developed and comparatively analyzed against a lightweight Convolutional Neural Network (CNN) architecture custom-designed for resource-constrained edge computing devices. While the ResNet152 V2 model achieved the highest accuracy (92.4%) in laboratory tests, the custom-designed CNN model proved to be the most practical and effective solution in real-world field tests conducted with a Raspberry Pi-based prototype, delivering 82.5% accuracy, a much smaller model size (31 MB), and superior processing speed (7.5 FPS). These results strongly demonstrate that application-specific, lightweight models can exhibit more robust and efficient performance in real-world scenarios compared to complex and large transfer learning architectures. The developed system has the potential to make significant contributions to agricultural sustainability and food security by optimizing pesticide use.