Self-supervised maize kernel classification and segmentation for embryo identification.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Deep learning (DL) in plant science aids food security and sustainability but requires extensive manual data annotation.
• Self-supervised learning (SSL) offers a solution by pre-training DL models on unlabeled datasets, reducing annotation burden.

Purpose of the Study:

• To implement and evaluate SSL methods (NNCLR, SimCLR) for maize kernel embryo classification and segmentation.
• To assess the annotation efficiency and transferability of SSL-pretrained models in agricultural applications.

Main Methods:

• Applied Nearest Neighbor Contrastive Learning of Visual Representations (NNCLR) and Simple Framework for Contrastive Learning of Visual Representations (SimCLR) for maize embryo analysis.
• Utilized high-throughput imaging data of maize kernels for classification and segmentation tasks.

Main Results:

• SSL techniques outperformed supervised transfer learning methods in both classification and segmentation.
• SSL models achieved competitive performance with as little as 1% annotated data, demonstrating high annotation efficiency.
• A single SSL pre-trained model was effectively fine-tuned for both classification and segmentation tasks, showing good transferability.

Conclusions:

• Self-supervised learning represents a significant advancement in data efficiency for agricultural deep learning and computer vision.
• SSL methods reduce the need for extensive manual annotation, making advanced image analysis more accessible in agriculture.