Data-efficient and accurate rapeseed leaf area estimation by self-supervised vision transformer for germplasms early...

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Breeding

Background:

• Accurate leaf area estimation is crucial for early-stage, high-throughput phenotyping in rapeseed breeding.
• Existing methods face challenges with expensive data annotation and leaf occlusion.
• Developing data-efficient solutions is essential for advancing agricultural technologies.

Purpose of the Study:

• To present a data-efficient deep learning framework for rapeseed leaf area quantification using smartphone RGB images.
• To address limitations of data annotation costs and leaf occlusion in plant phenotyping.
• To enable accurate and scalable non-destructive phenotyping for accelerated crop breeding.

Main Methods:

• A two-stage deep learning strategy utilizing a Vision Transformer (ViT) backbone.
• Self-supervised pre-training (DINOv2) on diverse public plant datasets, followed by fine-tuning on rapeseed data.
• Novel Canopy-Mix data augmentation and a hybrid loss function (Smooth L1, Log-Cosh) for occlusion handling and robust convergence.

Main Results:

• Achieved strong predictive performance with a Coefficient of Determination (R²)=0.805 via 5-fold cross-validation.
• Demonstrated high correlation between predicted leaf area and fresh weight (r=0.900) and dry weight (r=0.885).
• Outperformed baseline models, including those trained from scratch and pre-trained on ImageNet.

Conclusions:

• Domain-specific self-supervised pre-training effectively overcomes data limitations in agricultural vision.
• The proposed framework provides a robust and scalable tool for non-destructive phenotyping.
• This approach has the potential to significantly accelerate the rapeseed breeding cycle.