Transfer Learning and Fine-Tuned Faster R-CNN for Improved Insect Detection in Agriculture

Vesakkar Ansari, Jeseema Farhath (2024) Transfer Learning and Fine-Tuned Faster R-CNN for Improved Insect Detection in Agriculture. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (626kB) | Preview

Abstract

Over the years, various insect pests have posed challenges to the agricultural sector with serious off-takers to the losses. Correct identification of insects and pests are important steps in pest control, while existing solutions for this problem can be imprecise and inhibit scalability. Traditional methodologies are gradually losing its effective role in terms of identification of insects due to its incapability in processing large amount, and versatility of data and real time detection. To this end, this research seeks to apply the advanced deep learning method to improve insect detection in agricultural environments where the pest issue is prevalent. In particular, the examined architecture is based on the Faster R-CNN model, which follows the transfer learning approach where the base networks are trained on the pre-collected datasets, and then adapted to the authors’ custom collection of dangerous farm insects sourced on Kaggle. Various species of insects and temperature conditions are incorporated in this dataset making it rich for any training and testing of models. The primary innovation of this study lies in the development of a custom training pipeline that incorporates detailed accuracy calculations tailored for object detection tasks. This approach ensures the evaluation metrics accurately reflect the model's performance in detecting and localizing insects. The methodology also involves significant data augmentation to address the class imbalance inherent in the dataset, thereby improving the model's generalizability and robustness. Upon implementation, the fine-tuned Faster R-CNN model achieved a detection accuracy of 91%, demonstrating significant improvements compared to baseline models such as ResNet50V2, ResNet152V2, MobileNetV2, Xception which achieved accuracies of 72%, 63%, 70% and 53% respectively. Also after hyperparameter tuning efficiently, the best baseline model emerged to be the Xception model with an impressive accuracy of 78% on the validation data. These results highlight the superior performance of the Faster R-CNN and the Xception model in real-time pest monitoring and management. This enhanced detection capability can lead to more targeted pest control interventions, thereby reducing pesticide usage and promoting sustainable farming practices. This research contributes to the field of agricultural technology by providing a scalable and efficient solution for insect detection.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Milosavljevic, Vladimir UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science S Agriculture > S Agriculture (General)
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	26 Aug 2025 12:13
Last Modified:	26 Aug 2025 12:13
URI:	https://norma.ncirl.ie/id/eprint/8647

Actions (login required)

View Item