Synthetic data at scale: a development model to efficiently leverage machine learning in agriculture.

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Artificial intelligence (AI) and machine learning (ML) show great promise in agriculture.
• A major challenge for ML is the need for large, diverse datasets, which are expensive to collect and annotate.
• Existing methods for synthetic data generation have limitations.

Purpose of the Study:

• To present a novel, iterative model for cost-efficient synthetic training data generation.
• To demonstrate the model's application in developing an early disease detector for tomato plants.
• To offer a quantitative approach for refining synthetic data generation, avoiding subjective human assessment.

Main Methods:

• Development of an iterative model for synthetic data generation.
• Training a neural classifier exclusively on synthetically generated images.
• Iterative refinement of the synthetic image generation process.
• Quantitative evaluation of the model's performance and cost-efficiency.

Main Results:

• The developed early disease detector for tomato plants achieved superior generalization performance.
• The approach demonstrated higher cost-efficiency compared to traditional synthetic data generation methods.
• The model successfully trained a neural classifier using only synthetic images.

Conclusions:

• The proposed iterative model offers a cost-effective solution for generating synthetic training data in agriculture.
• This method can significantly reduce the development costs associated with ML applications in agriculture.
• The quantitative approach enhances the reliability and performance of ML models trained on synthetic data.