Machine Learning Approaches for Rice Seedling Growth Stages Detection.

Area of Science:

• Agricultural Science
• Computer Science
• Remote Sensing

Background:

• Timely field operations in rice cultivation depend on accurate seedling growth stage recognition.
• Manual visual inspection is labor-intensive, subjective, and inefficient for large-scale agriculture.
• Machine learning on Unmanned Aerial Vehicle (UAV) imagery offers a high-throughput, non-invasive alternative for phenotyping.

Purpose of the Study:

• To develop and evaluate automatic approaches for detecting three critical rice seedling growth stages (BBCH11, BBCH12, BBCH13).
• To compare the performance of traditional machine learning algorithms with deep learning models for this classification task.

Main Methods:

• UAV images were captured vertically at a 3-m height.
• A dataset was created with images of three rice growth stages, three cultivars, five densities, and varying sowing dates.
• Traditional methods used Histograms of Oriented Gradients (HOG) with Support Vector Machine (SVM).
• Deep learning approaches employed state-of-the-art Convolutional Neural Network (CNN) models, including Efficientnet, VGG16, Resnet50, and Densenet121.

Main Results:

• The best HOG-SVM model achieved 84.9% accuracy, 85.9% average precision, 84.9% average recall, and 85.4% F1 score.
• EfficientnetB4 demonstrated superior performance among deep learning models, reaching 99.47% accuracy, 99.53% average precision, 99.39% average recall, and 99.46% F1 score.
• Deep learning models significantly outperformed traditional machine learning methods.

Conclusions:

• Automatic detection of rice seedling growth stages using machine learning on UAV imagery is highly effective and efficient.
• Deep learning, specifically EfficientnetB4, provides a robust solution for accurate, large-scale rice phenotyping and crop management.
• The proposed method can serve as a valuable tool for optimizing agricultural practices and increasing yield.