SoyCountNet: a deep learning framework for counting and locating soybean seeds in field environment

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate soybean seed counting and spatial localization (Seeds Per Plant - SPP) are crucial for yield estimation and cultivar evaluation.
• Field conditions present challenges like complex backgrounds, pod occlusion, and uneven grain filling, hindering traditional counting methods.

Purpose of the Study:

• To develop SoyCountNet, a deep learning framework for automatic soybean seed counting and localization at the single-plant level in field conditions.
• To enhance the accuracy and robustness of soybean seed quantification for agricultural applications.

Main Methods:

• Utilized a field-based phenotyping platform and optimized the Point-to-Point Network (P2PNet).
• Employed a VGG19_BN backbone and Super Token Sampling Vision Transformer (SViT) for feature extraction.
• Integrated Efficient Channel Attention (ECA) for feature fusion and an improved loss function for enhanced precision and spatial consistency.

Main Results:

• SoyCountNet outperformed existing methods on a field soybean dataset.
• Achieved a Mean Absolute Error (MAE) of 4.61, Root Mean Square Error (RMSE) of 6.03, and R² of 0.94.
• Demonstrated consistent performance across different soybean cultivars for reliable SPP estimates.

Conclusions:

• SoyCountNet provides a reliable and scalable solution for precise soybean seed counting and localization in complex field environments.
• The lightweight architecture facilitates deployment on intelligent agricultural platforms for high-throughput phenotyping and precision breeding.
• This work supports the advancement of intelligent and sustainable agricultural technologies.