Training instance segmentation neural network with synthetic datasets for crop seed phenotyping.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Manual data annotation is a bottleneck in training neural networks for plant phenotyping.
• Developing automated methods for plant trait analysis is crucial for agricultural advancements.

Purpose of the Study:

• To investigate the efficacy of purely synthetic data for training instance segmentation neural networks for plant phenotyping.
• To reduce the time and cost associated with manual data annotation in agricultural machine learning applications.

Main Methods:

• Utilized domain randomization to generate a large dataset of synthetic barley seed images with random orientations.
• Trained an instance segmentation neural network exclusively on this synthetic dataset.
• Validated the model's performance on a real-world test dataset.

Main Results:

• The neural network trained on synthetic data achieved high performance metrics: 96% recall and 95% average precision on real-world barley seed images.
• The approach demonstrated effectiveness across various crops, including rice, lettuce, oat, and wheat.
• Synthetic data generation significantly reduced the need for manual annotation.

Conclusions:

• Purely synthetic data, generated using domain randomization, is sufficient for training effective plant phenotyping models.
• This method offers a scalable and cost-effective solution to accelerate the deployment of deep learning in agriculture.
• Automated data generation can overcome limitations in manual data annotation for agricultural AI.