Abhinav Pagadala: Precision mechanical robotic weeding for specialty crops

Bionote

Abhinav Pagadala is a M.S. student in Statistics, Applied Analytics at the University of Illinois Urbana-Champaign. He graduated from Kansas State University with a B.S. in Computer Engineering. His current position is as a Graduate Research Assistant in the Agricultural & Biological Engineering Department where his focus is computer vision and AI-driven decision-making for specialty crops in precision and digital agriculture. His recent work involves developing object detection and tracking models to support an AI-based weed management system for Illinois horseradish farmers. Abhinav is also preparing a manuscript comparing new YOLO technologies for agricultural applications leveraging analytics and data-driven decision-making. He is passionate about solving real-world challenges through data science and machine learning.

Within horseradish production, one of the critical challenges farmers face is effective weed management. With the lack of mechanization available, farmers rely heavily on manual labor which is often inefficient. This study aims to compare emerging computer vision technologies as a part of an AI-based weed management system. As the world of precision and digital agriculture becomes more relevant, previous research has not had the opportunity to utilize state of the art You Only Look Once (YOLO) models for agricultural applications. We use data collected from commercial growers in Illinois to train, test, and optimize various custom YOLO models, considering a wide range of parameters to determine the ideal choice. With YOLOv12 being the newest release of You Only Look Once (YOLO) models, it has significant architectural changes from previous iterations, causing implications on training and overall efficiency. Our findings support the use of YOLO11 instead of the newest YOLOv12, when taking evaluation criteria regarding accuracy (YOLO11 is a ~3% improvement), likewise for inference speeds, training time, and computational efficiency. These results highlight the importance of tailoring model selection for domain-specific and system constraints.

Co-Researcher: Sunoj Shajahan, John Reid, Elizabeth Wahle, Dennis Bowman, Pavan Dabilpuram, Janmejay Rathi, Sandesh Poudel