Data-Efficient and Accurate Rapeseed Leaf Area Estimation by Self-supervised Vision Transformer for Germplasms Early Evaluation

Early-stage, accurate and high-throughput phenotyping‌ through leaf area estimation is ‌critical‌ for future rapeseed breeding, but faces ‌two key constraints‌: expensive data annotation and persistent challenge of leaf occlusion. To address these issues, we present a ‌data-efficient‌ deep learning framework using smartphone-captured top-down RGB images for rapeseed leaf area quantification. Our approach utilizes a two-stage strategy where a Vision Transformer (ViT) backbone is first pre-trained on a large, aggregated corpus of diverse, non-rapeseed public plant datasets using the DINOv2 self-supervised learning method. This pre-trained model is then fine-tuned on a custom rapeseed dataset using a novel Canopy-Mix data augmentation technique to handle fragmented views analogous to occlusion, and a hybrid loss function combining Smooth L1 and Log-Cosh for robust convergence. Through rigorous 5-fold cross-validation, our proposed model achieved state-of-the-art predictive performance (Coefficient of Determination, R$^2$=0.805). What’s more, the predicted leaf area demonstrated a remarkably strong correlation with both fresh weight (r=0.900) and dry weight (r=0.885). The model significantly outperformed a range of baselines, including models trained from scratch, those pre-trained on ImageNet, and a heuristic method based on manually annotated bounding boxes. Ablation studies confirmed the essential contribution of each component, while qualitative analysis of attention maps demonstrated the model's ability to precisely localize the leaf canopy and ignore background distractors. This study demonstrates that domain-specific self-supervised pre-training offers a powerful solution to overcome data limitations in agricultural vision, providing a robust and scalable tool for non-destructive phenotyping that can potentially accelerate the rapeseed breeding cycle.