By integrating a lightweight dilated contextual adapter (DCon-Adapter) and a weight decomposition matrix (WDM), the model learns efficiently from limited samples, achieving 93.53% accuracy in controlled tests and outperforming existing approaches in real-world settings.
Plant disease recognition technologies have advanced rapidly thanks to deep learning and large annotated datasets, but agricultural applications face unique hurdles. Data collection in the field is expensive and time-consuming, and some diseases are so rare that acquiring sufficient samples is nearly impossible. Few-shot learning offers a solution, enabling models to learn from just a few labeled examples per class. Yet, conventional methods often require pretraining on large, domain-specific datasets—a resource rarely available in agriculture. Foundation models, such as CLIP and DINO, have shown strong performance in zero- and few-shot learning, but their generalization to agricultural imagery is limited by domain differences and class imbalances.
A study (DOI: 10.1016/j.plaphe.2025.100024) published in Plant Phenomics on 28 February 2025 by Ruifang Zhai ’s team, Huazhong Agricultural University, improves plant disease recognition accuracy and generalization in data-limited scenarios, offering a practical solution for real-world agricultural diagnostics.
The researchers implemented PlantCaFo, a few-shot plant disease recognition model, by leveraging pretrained backbone networks from foundation models—CLIP (ResNet-50 image encoder and Transformer text encoder), DINO (ResNet-50), and DINO2 (distilled ViT-S/14). Training was conducted with varying sample sizes (1, 2, 4, 8, and 16 shots) using consistent random seeds. Only the cache model, dilated contextual adapter (DCon-Adapter), and weight decomposition matrix (WDM) were trainable, optimizing efficiency. PlantCaFo and its enhanced variant PlantCaFo* (with Mixup and CutMix augmentations) were trained for 40 epochs using AdamW, with evaluation on fixed-size test sets. Experiments on the PlantVillage dataset revealed that while Tip-Adapter-F performed well in ultra-low-shot settings (2–4 shots), PlantCaFo and PlantCaFo* surpassed it in higher-shot scenarios, outperforming CaFo-Base by up to 4.60% and achieving consistent gains on the more challenging Cassava dataset. Confusion matrices confirmed high classification accuracy and minimal misclassifications. Although runtime on PlantVillage doubled relative to CaFo-Base due to larger data handling, accuracy gains of up to 7.74% justified the trade-off. Generalization tests on an out-of-distribution dataset (PDL) showed strong performance on split1 (single-species diseases) but reduced accuracy on split2 (multi-species diseases with complex backgrounds), indicating domain shift challenges. Ablation studies demonstrated that the DCon-Adapter contributed more to performance than the WDM, with their combination yielding further gains, particularly when coupled with data augmentation. Prompt-based experiments confirmed PlantCaFo’s superior text–image understanding even with simple templates. Visualizations using Smooth Grad CAM++ revealed that, compared to CaFo-Base, PlantCaFo more effectively focused on disease-relevant regions while filtering irrelevant features, albeit with slightly less precise localization due to its broader generalization across species. These results highlight PlantCaFo’s capacity to balance accuracy, efficiency, and adaptability for diverse plant disease identification tasks under data-scarce conditions.
PlantCaFo’s ability to accurately recognize plant diseases from minimal data could transform agricultural diagnostics, particularly in resource-limited settings. Farmers, agronomists, and plant health agencies could rapidly deploy AI-based disease detection tools without the prohibitive costs of collecting large training datasets. This efficiency makes the technology suitable for mobile apps, drone-based monitoring systems, and early-warning platforms that help curb disease spread and reduce crop losses.
##
References
DOI
10.1016/j.plaphe.2025.100024
Original Source URL
https://doi.org/10.1016/j.plaphe.2025.100024
Funding information
This work was supported by the National Key Research and Development Program of China (2023YFF1000100).
About Plant Phenomics
Science Partner Journal Plant Phenomics is an online-only Open Access journal published in affiliation with the State Key Laboratory of Crop Genetics & Germplasm Enhancement, Nanjing Agricultural University (NAU) and distributed by the American Association for the Advancement of Science (AAAS). Like all partners participating in the Science Partner Journal program, Plant Phenomics is editorially independent from the Science family of journals. Editorial decisions and scientific activities pursued by the journal's Editorial Board are made independently, based on scientific merit and adhering to the highest standards for accurate and ethical promotion of science. These decisions and activities are in no way influenced by the financial support of NAU, NAU administration, or any other institutions and sponsors. The Editorial Board is solely responsible for all content published in the journal. To learn more about the Science Partner Journal program, visit the SPJ program homepage.