Maize Silage Kernel Fragment Estimation Using Deep Learning-Based Object Recognition in Non-Separated Kernel/Stover...

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate corn silage kernel processing evaluation is crucial for crop quality assessment.
• Current evaluation methods are time-consuming, requiring manual separation of kernels and stover.

Purpose of the Study:

• To introduce and evaluate deep learning methods for predicting corn silage kernel processing quality.
• To enable kernel detection and processing score prediction directly from images without manual sample preparation.

Main Methods:

• Two deep learning models (bounding-box and instance segmentation) were developed and trained on annotated images of corn silage kernels.
• Networks were trained on 1393 images with over 6907 manually annotated kernel instances.
• Model performance was evaluated using average precision and correlation analysis with manual Kernel Processing Score (KPS) annotations.

Main Results:

• Instance segmentation achieved an average precision of 36.1% at an IoU of 0.5, while bounding-box detection achieved 34.0%.
• A strong correlation (r = 0.88, p < 0.0001) was found between predicted and manually annotated Kernel Processing Scores (KPS).
• The models demonstrated robustness across images from three different harvest seasons.

Conclusions:

• Deep learning object recognition offers a rapid and automated approach for corn silage kernel processing evaluation.
• This technology has the potential to reduce quality assessment time from hours/days to minutes.
• The developed methods can significantly assist farmers in making timely decisions during the harvesting season.