Benchmarking Self-Supervised Contrastive Learning Methods for Image-Based Plant Phenotyping.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Self-supervised learning (SSL) offers a way to use unlabeled plant phenotyping data.
• SSL applications in plant phenotyping, especially for detection and counting, are under-explored.

Purpose of the Study:

• Benchmark SSL methods (MoCo v2, DenseCL) against supervised learning for plant phenotyping tasks.
• Evaluate the impact of pretraining dataset domain and redundancy on SSL performance.
• Analyze learned representations from different pretraining strategies.

Main Methods:

• Compared momentum contrast (MoCo) v2 and dense contrastive learning (DenseCL) with supervised learning.
• Transferred learned representations to wheat head detection, plant instance detection, wheat spikelet counting, and leaf counting.
• Investigated effects of source dataset domain and redundancy on downstream task performance.

Main Results:

• Supervised pretraining generally outperformed SSL methods.
• MoCo v2 and DenseCL learned distinct representations compared to supervised methods.
• Diverse, domain-similar pretraining datasets maximized downstream performance.
• SSL methods appeared more sensitive to pretraining data redundancy than supervised methods.

Conclusions:

• SSL methods require careful consideration of pretraining data for effective plant phenotyping.
• Supervised learning remains a strong baseline for image-based plant phenotyping tasks.
• This study provides guidance for developing improved SSL techniques in this domain.