Recognition of soybean pods and yield prediction based on improved deep learning model

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Soybean pod determination is crucial for yield and quality.
• Accurate pod counting and weight estimation are vital for soybean production.

Purpose of the Study:

• To develop an improved YOLOv5 model for soybean pod recognition and weight estimation.
• To enhance the accuracy and efficiency of soybean yield prediction.

Main Methods:

• Improved YOLOv5 model incorporating coordinate attention (CA) and boundary box regression loss for pod detection and counting.
• Back propagation (BP) neural network for estimating soybean plant yield based on pod data.
• Comparative analysis against the traditional YOLOv5 model.

Main Results:

• The improved YOLOv5 model achieved 91.7% average precision (AP) and a detection rate of 24.39 frames per millisecond.
• The model demonstrated high accuracy in estimating single pod weight (MSE=0.00865, R²=0.945) and total plant yield (MRE=0.122).
• The proposed model reduced calculation and parameters by 17% and 7.6% compared to the traditional YOLOv5.

Conclusions:

• The developed method provides a robust and efficient solution for real-time soybean pod detection and yield estimation.
• This technology supports intelligent breeding and precision agriculture in soybean cultivation.