Detecting common coccinellids found in sorghum using deep learning models.

Area of Science:

• Agricultural Science
• Entomology
• Computer Science

Background:

• Sorghum production faces challenges from pests like the sugarcane aphid (SCA), necessitating efficient management strategies.
• Current field scouting for SCA and its natural enemies is labor-intensive and costly, impacting integrated pest management (IPM).
• Natural enemies, primarily coccinellids, are crucial for biological control of SCA, but their detection is inefficient.

Purpose of the Study:

• To develop and train machine learning models for automated detection and classification of key coccinellid species in sorghum.
• To establish deep learning software for enhancing IPM by facilitating the identification of natural enemies.

Main Methods:

• Training object detection models, including Faster Region-based Convolutional Neural Network (Faster R-CNN) with Feature Pyramid Network (FPN), YOLOv5, and YOLOv7.
• Utilizing a dataset of coccinellid images sourced from the iNaturalist project for model training and evaluation.
• Evaluating model performance using standard object detection metrics like average precision (AP) and AP@0.50.

Main Results:

• The YOLOv7 model demonstrated superior performance in detecting and classifying coccinellids, achieving an AP@0.50 of 97.3% and an AP of 74.6%.
• Faster R-CNN-FPN, YOLOv5, and YOLOv7 models were successfully trained to identify seven common coccinellid species found in sorghum.
• The developed models provide a foundation for automated insect detection in agricultural settings.

Conclusions:

• Automated deep learning software offers a viable solution for efficient detection of natural enemies in sorghum fields.
• The YOLOv7 model shows significant promise for practical application in integrated pest management for sorghum.
• This technology can reduce reliance on chemical insecticides by promoting conservation of natural predators.