A preprint on arXiv introduces A100, a vision foundation model designed for agricultural plant images. The authors pretrain a hybrid CNN-transformer backbone on 3 million unlabeled field images covering more than 100 crop species, combining MBConv stages for local features with transformer stages for global context. Pretraining uses DINO self-distillation, iBOT masked patch reconstruction, and KoLeo uniformity regularization.
Without fine-tuning, the frozen backbone outperforms supervised ImageNet-12K representations on agricultural tasks: 95.37% accuracy on crop classification, a 2.56-point gain, and a 10.4% lower leaf-count error on the GroMo25 phenotyping benchmark. After full fine-tuning, A100 reaches 99.90% on 32-class crop identification and 99.02% on 44-class disease detection, and it beats the CvT-13 baseline on the PlantDoc dataset.
The authors emphasize label efficiency: with just 1% of labeled data, roughly 284 images per category, A100 achieves 98.01% accuracy on plant classification. Under the same protocol, it outperforms the publicly released agricultural models FoMo4Wheat and LeafVision by more than 14 points. The team plans to release the pretrained model publicly, which could lower the annotation barrier for crop monitoring and disease diagnosis in data-scarce regions.