Geographic-Scale Coffee Cherry Counting with Smartphones and Deep Learning

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Traditional plant monitoring methods like remote sensing and drones are often infeasible for crops under tree canopies, such as coffee.
• This limitation hinders large-scale, cost-effective plant monitoring and phenotyping for coffee production.

Purpose of the Study:

• To develop a geographic-scale coffee cherry counting method using an AI-powered citizen science approach.
• To enable scalable and low-cost phenotyping for coffee crops, overcoming limitations of existing technologies.

Main Methods:

• Utilized basic smartphones for image capture of coffee trees, involving nearly 1,000 smallholder farmers across Peru and Colombia.
• Trained and validated the YOLO (You Only Look Once) v8 model for cherry detection using 8,904 images from 2,968 trees.
• Estimated total cherries per tree by multiplying the average cherries per picture by the number of branches.

Main Results:

• The AI model achieved an R² of 0.59 in Peru and 0.71 when tested in Colombia, demonstrating transferability across varieties and conditions.
• Overall performance across both countries reached an R² of 0.72.
• This marks the first AI-powered method for coffee cherry counting, achieving geographic-scale monitoring.

Conclusions:

• The developed AI citizen science method is a scalable and transferable solution for coffee cherry counting and phenotyping.
• This approach has the potential for multiyear, photo-based monitoring in low-income countries worldwide.
• It offers a cost-effective alternative for coffee crop management and research.