Convolutional Neural Net-Based Cassava Storage Root Counting Using Real and Synthetic Images

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Cassava storage root number and size are key indicators of crop yield and quality.
• Manual and semi-automatic counting methods are labor-intensive and prone to errors due to root occlusion.
• Limited annotated data hinders the development of Convolutional Neural Network (CNN)-based plant feature counting methods.

Purpose of the Study:

• To develop an automated system for accurately counting cassava storage roots from images.
• To address the challenge of limited annotated data for CNN models in plant phenotyping.
• To enable direct image-to-count prediction for cassava root analysis.

Main Methods:

• Utilized a conditional Generative Adversarial Network (GAN) to generate synthetic cassava root images for data augmentation.
• Developed a direct image-to-count prediction model trained on both real and synthetic images.
• Implemented a system that first predicts root age group (young/old) and then applies an age-specific counting model.

Main Results:

• Achieved 91% accuracy in predicting the age group of cassava storage roots.
• Demonstrated 86% and 71% overall percentage agreement in counting 'old' and 'young' storage roots, respectively.
• Successfully showed that synthetic data can supplement missing classes for direct counting tasks.

Conclusions:

• Synthetically generated images effectively augment training datasets for plant feature counting.
• The developed system provides an accurate and automated solution for counting cassava storage roots.
• This approach advances the application of deep learning in agricultural phenotyping and yield estimation.