Researchers at the LSU AgCenter are using computer vision and artificial intelligence to improve the way soybean seed quality is evaluated which they say could eliminate the need for time-intensive visual inspections.

Using high-resolution 2D images of soybean seed surfaces, researchers trained AI models to identify common defects that can affect seed quality. Types of defects include wrinkles, dark spots and other types of damage.

“Accurate seed quality assessment is critical for soybean producers, seed processors and researchers,” said Ivan Grijalva, assistant professor in the entomology department. “Our goal was to develop a model that can evaluate seed damage quickly and reduce the subjectivity of manual inspections.”

Thousands of seed images were collected and annotated. Researchers labeled visible damage on individual seeds, creating a dataset that trained the vision models to recognize and classify damage to the seeds.

According to Grijalva, building a high-quality training dataset was one of the most important steps in the process.

“Computer vision models require high-quality data to train them accurately,” he said. “By carefully annotating seed images and documenting different types of damage, we were able to create a strong foundation for developing reliable detection models.”

The team evaluated several computer vision approaches to determine which methods performed best. YOLOv8 and YOLO11 models were used for seed detection and counting. A U-Net model with a ResNet-34 backbone was employed for more detailed segmentation of damaged areas on seed surfaces.

Researchers also applied image preprocessing techniques, including image rotation and brightness adjustments, to increase variation within the dataset and improve model performance.

Grijalva helped guide model development and data analysis efforts, ensuring the team evaluated performance using standard performance metrics such as precision, recall and F1-score.

“We wanted to identify models that could perform reliably under practical conditions,” Grijalva said. “The evaluation process allowed us to find the best approach that balanced accuracy and consistency.”

In testing, the selected model achieved approximately 82% accuracy in detecting and classifying seed damage.

The results demonstrate the growing potential of artificial intelligence tools in agricultural quality assessment, according to Grijalva.

The research team plans to incorporate the model into a decision-support tool that could eventually be used in automated seed grading systems.

Soybean seed images with different surface damages were used to train AI models. Various defects include: A) undamaged seed, B) wrinkle with dark spot damages, C-F) multiple damages, G-H) dark spot damages.

Detection of different damages using a computer vision model. Photos by Peyton Falterman