Identifying Cocoa Flower Visitors: A Deep Learning Dataset.

Area of Science:

• Agricultural Science
• Computer Vision
• Entomology

Background:

• Cocoa production is a significant global industry, yet research on yield enhancement via pollination is limited.
• Advancements in embedded hardware and AI enable detailed analysis of cocoa flower visitors and their impact on yields.

Purpose of the Study:

• To present the first comprehensive dataset of cocoa flower visitors, including various insect families and background images.
• To evaluate the performance of different YOLOv8 deep learning models for identifying insects in cocoa plantations.
• To establish a benchmark for deep learning model performance on low-contrast images with challenging detection targets.

Main Methods:

• Curated a dataset of 5,792 insect images (Ceratopogonidae, Formicidae, Aphididae, Araneae, Encyrtidae) and 1,082 background images from 23 million images collected over two years.
• Utilized embedded cameras in Hainan province, China, for image acquisition.
• Trained and tested various sizes of YOLOv8 models, progressively increasing the background image ratio in the training set.

Main Results:

• The medium-sized YOLOv8 model demonstrated the best performance with an 8% background image ratio, achieving an F1 Score of 0.71 and mAP50 of 0.70.
• The dataset proved effective for comparing deep learning model architectures on challenging image datasets.
• Identified optimal model configuration for accurate insect detection in complex cocoa plantation environments.

Conclusions:

• The presented dataset is valuable for advancing research in cocoa pollination monitoring and sustainable agriculture.
• Deep learning models, particularly YOLOv8, can effectively identify cocoa flower visitors, contributing to yield improvement strategies.
• This work supports future efforts in precision agriculture and automated crop management through AI-driven data analysis.