Deep learning-based accurate detection of insects and damage in cruciferous crops using YOLOv5

Abstract

• Automatic identification system of important insects associated with cruciferous crops and some of their damage symptoms using deep learning-based state-of-the-art single-stage object detection models. • Evaluation and comparison of performance metrics, loss functions, computational costs, and inference speeds of five variants of YOLOv5 viz. nano (n), small (s), medium (m), large (l), and extra-large (x). • YOLOv5l performed better than other variants with an outstanding testing accuracy of 99.5%. • The performances of YOLOv5l were compared to newer YOLO models viz . YOLOv7, YOLOv8, YOLOv9, YOLOv10, and YOLOv11. • For the first time, AI-based insect identification is not only confined to the pestiferous species, but also covers the beneficial insects. Insects are an integral part of an agroecosystem. Some of them are pestiferous, while some are beneficial like- natural enemies and pollinators. Therefore, it is very important to identify and manage them timely. With the rapid development of convolutional neural networks, automatic detection techniques for identifying insects using digital images have shown impressive performances in agriculture. In this study, we propose a deep learning approach using YOLOv5-based single-stage object detection model for the identification of agriculturally important insects of crucifers and some of their damage symptoms. A total of 2,730 images were captured from different fields and polyhouses using different smartphones and an SLR camera. The specimens were taxonomically identified by experts and the images were curated, annotated, resized, augmented, split, and trained, validated and tested through five variants of YOLOv5 viz. nano (n), small (s), medium (m), large (l), and extra-large (x). After all the experiments, YOLOv5l was found to be the best-performing model, acquiring an average accuracy, precision, recall, and F1-Score of 99.5%, 92.0%, 83.0%, and 0.873, respectively in the test images. The inference time and computational complexity of YOLOv5l are also significantly lower than those of YOLOv5x. Therefore, to strike a balance between complexity and performance, YOLOv5l has emerged as the most viable option to integrate with AI-based insect identification applications. Our findings reveals that deep learning is reliable for quick detection of insects under complex backgrounds. Further, we demonstrate that use of damage symptoms produced by insects will also be explored for pest detection. Integration of present model with mobile application will help the farmers and stake holders in detection of insects and suggesting effective management. This study presents a deep learning framework for the accurate identification of agriculturally important insects associated with cruciferous crops with some of their damage symptoms. Using image preprocessing and leveraging advanced convolutional neural networks, the model efficiently identifies pest species, and their presence with specific crop damage indicators. The pollinators and predatory insects are also detected. Extensive field data, rigorous validation and testing demonstrate the model's effectiveness and reliability, promising substantial improvements in crop protection and yield optimization.

Topics & Concepts