Comparison of Different Machine Learning Algorithms for the Prediction of the Wheat Grain Filling Stage Using RGB...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Wheat grain filling is crucial for yield but vulnerable to environmental stresses like high temperatures.
• Monitoring wheat grain development is essential for timely agricultural interventions and yield optimization.
• Existing methods for tracking grain development lack quantitative precision from visual data.

Purpose of the Study:

• To develop a comprehensive dataset (WheatGrain) for analyzing wheat grain filling dynamics.
• To quantitatively predict days after anthesis (DAA) using RGB images of wheat grains.
• To compare the efficacy of traditional machine learning, deep learning, and few-shot learning for DAA prediction.

Main Methods:

• Creation of the WheatGrain dataset, comprising thousands of wheat grain RGB images from 6 to 39 DAA.
• Implementation and evaluation of Random Forest (RF), deep learning models (including Vision Transformer - ViT), and few-shot learning algorithms.
• Comparative analysis of model performance based on precision, recall, and accuracy for DAA prediction.

Main Results:

• Deep learning algorithms significantly outperformed traditional machine learning (Random Forest).
• Vision Transformer (ViT) achieved the highest precision (99.03%) and recall (99.00%) among deep learning models.
• Few-shot learning demonstrated high accuracy (96.86%) and recall (96.67%) in 5-shot scenarios, indicating its utility for fine-grained recognition with limited data.

Conclusions:

• The WheatGrain dataset enables prompt monitoring of wheat grain filling dynamics.
• ViT enhances deep learning performance for DAA prediction, while few-shot learning reduces data requirements.
• This approach offers a novel method for monitoring wheat development, aiding in disaster prevention and yield improvement.