Development of AI-Driven Pest Detection System Using Drone Imagery

Thesis Abstract

In recent years, pest management in agriculture has faced significant challenges due to the increasing prevalence of pest infestations that threaten crop yields and food security, coupled with the limitations of traditional surveillance methods which are often labor-intensive, time-consuming, and subject to human error. This study aims to develop an Artificial Intelligence (AI)-driven pest detection system leveraging drone imagery to facilitate rapid, accurate, and cost-effective identification of pest outbreaks in large-scale agricultural fields. The specific objectives include designing a drone-based imaging framework for capturing high-resolution data, developing and training machine learning models—specifically convolutional neural networks (CNNs)—for pest identification, evaluating the system's detection accuracy, and assessing its operational efficiency within real-world farm environments. The research adopts a quantitative, experimental design to evaluate the efficacy of the developed pest detection system. The study's population comprises maize and cotton farms covering an area of approximately 10,000 hectares across the southeastern region of the country, selected based on pest prevalence history and farm management practices. A stratified random sampling method is employed to select 50 farms, with a total of 200 drone flights conducted to acquire imagery data under varying environmental conditions and pest infestation levels. Data collection instruments include multispectral drone cameras capable of capturing visible and near-infrared imagery, and a custom-developed AI model trained using a labeled dataset of 10,000 annotated images derived from field sampling and expert pest identification. The collected imagery dataset is preprocessed to enhance feature extraction, followed by training of the CNN model using transfer learning on a subset of 8,000 images, with the remaining 2,000 images reserved for testing. Model performance is evaluated through accuracy metrics such as precision, recall, F1-score, and Receiver Operating Characteristic (ROC) curves, employing techniques like cross-validation to ensure robustness. Additionally, model interpretability is assessed via Grad-CAM visualizations, and the operational efficiency of the system is analyzed through time-to-detection and false-positive rates. It is anticipated that the AI-driven system will demonstrate a detection accuracy exceeding 85% in identifying major pests such as fall armyworm and cotton bollworm, with a significant reduction in detection time relative to manual scouting methods. The study expects to reveal that drone imagery combined with deep learning models can accurately discern pest-infested zones, facilitating targeted intervention and precision pest management. These findings are projected to contribute novel insights into the application of UAV-based AI tools for real-time pest surveillance, filling existing gaps in scalable, automated pest detection technologies, and providing a framework adaptable across various crop types and agro-ecological zones. The research's primary contribution to knowledge lies in integrating drone technology with cutting-edge AI algorithms to transform pest monitoring paradigms, promoting sustainable agricultural practices through targeted pest control, and reducing reliance on chemical interventions. The study concludes with recommendations for large-scale adoption of drone-based pest detection systems, emphasizing the need for policy frameworks to support technological deployment, ongoing system refinement through adaptive learning, and integration with existing farm management systems. It also advocates for further research on expanding the model’s applicability to multispecies pest detection, incorporating multispectral and hyperspectral sensing, and exploring user interfaces for farmer accessibility and real-time decision-making support.