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

## TL;DR

SoyCountNet is a deep learning model that accurately counts and locates soybean seeds in field conditions, improving yield estimation and cultivar evaluation.

## Contribution

SoyCountNet introduces a novel lightweight deep learning framework for soybean seed counting and localization in field environments.

## Key findings

- SoyCountNet achieves a mean absolute error (MAE) of 4.61 and a coefficient of determination (R²) of 0.94 on the field soybean dataset.
- The model provides reliable seed counts across different soybean cultivars in complex field conditions.
- SoyCountNet's lightweight design enables deployment on intelligent agricultural platforms for high-throughput phenotyping.

## Abstract

Accurate counting and spatial localization of soybean seeds—particularly Seeds Per Plant (SPP)—are critical for yield estimation and cultivar evaluation. In field environments, however, complex backgrounds, pod occlusion, and uneven grain filling make high-precision counting challenging, and traditional methods often struggle to balance accuracy and robustness.

To address these challenges, this study proposes SoyCountNet, a deep learning framework for automatic soybean seed counting and localization at the single-plant level under field conditions. The model is built on a self-constructed field-based phenotyping platform and optimized using the lightweight Point-to-Point Network (P2PNet). For feature extraction, a VGG19_BN backbone and a Super Token Sampling Vision Transformer (SViT) module are employed to enhance local feature representation and global contextual understanding. During feature fusion, the Efficient Channel Attention (ECA) mechanism strengthens seed-related features while suppressing interference from leaves, stems, and soil. Furthermore, an improved loss function that combines point-distance constraints with overlap penalties enhances both counting precision and spatial consistency.

Experimental results demonstrate that SoyCountNet outperforms existing approaches on the field soybean dataset. It achieves a mean absolute error (MAE) of 4.61, a root mean square error (RMSE) of 6.03, and a coefficient of determination (R²) of 0.94. The model demonstrates consistent performance across the tested soybean cultivars, providing reliable SPP estimates within the evaluated dataset.

These findings indicate that SoyCountNet offers a reliable and scalable solution for precise soybean seed counting and localization in complex field environments. Its lightweight architecture allows deployment on intelligent agricultural platforms, supporting high-throughput phenotyping, yield prediction, and precision breeding, while providing a foundation for the future development of intelligent and sustainable agricultural technologies.