Soybean leaf estimation based on RGB images and machine learning methods.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• RGB photographs offer a dynamic method for estimating crop growth, crucial for understanding plant physiology.
• Traditional leaf parameter measurements are laborious and time-consuming, hindering efficient crop breeding.
• Accurate estimation of soybean leaf parameters is vital for accelerating breeding programs.

Purpose of the Study:

• To develop and evaluate machine learning models for precise soybean leaf parameter estimation using RGB images.
• To compare the performance of different regression models for predicting leaf number, fresh weight, and area index.
• To introduce a novel, efficient technique for soybean phenotyping.

Main Methods:

• Image segmentation using a Unet neural network to isolate soybean leaves from RGB images.
• Development and comparison of Random Forest, Cat Boost, and Simple Nonlinear Regression models for parameter estimation.
• Validation of model performance using metrics such as Intersection over Union (IOU), PA, Recall, and Average Testing Prediction Accuracy (ATPA).

Main Results:

• Unet achieved high soybean image segmentation accuracy with IOU, PA, and Recall values of 0.98, 0.99, and 0.98, respectively.
• Random Forest demonstrated superior performance over Cat Boost and Simple Nonlinear Regression in predicting leaf number, fresh weight, and leaf area index.
• Random Forest models achieved ATPAs of 73.45% (leaf number), 74.96% (leaf fresh weight), and 85.09% (leaf area index).

Conclusions:

• The Unet neural network accurately segments soybeans from RGB images, facilitating subsequent analysis.
• The Random Forest model exhibits strong generalization capabilities and high accuracy for estimating key soybean leaf parameters.
• Integrating advanced machine learning with digital imaging provides an effective approach for enhancing soybean phenotyping.