Huazhong Agricultural University has made new progress in plant disease detection under limited data conditions. A research team led by associate professor Zhai Ruifang from the College of Informatics recently published a paper titled "PlantCaFo: An efficient few-shot plant disease recognition method based on foundation models" in Plant Phenomics, proposing a lightweight and high-performing recognition model — PlantCaFo.

Accurate disease identification is key to safeguarding food security. However, most current artificial intelligence models rely on large volumes of annotated images, limiting their real-world applicability in complex agricultural environments. Addressing the challenge of scarce field data, the team developed PlantCaFo, a simplified CaFo-based model that leverages generative AI (such as GPT) to build high-quality image-text datasets. It integrates vision-language foundation models such as CLIP and DINO, enhancing performance even with minimal samples.

The model introduces two innovations — the DCon-Adapter and Weight Decomposition Module (WDM) — which improve the model's ability to detect subtle disease traits on both the global and local scale.

Tests showed that with only four images per class, PlantCaFo achieved 82.63 percent accuracy on the Plant Village dataset—4.6 percent higher than the baseline. With 16 images, it reached 93.53 percent, outperforming existing models. On the more complex Cassava dataset, it improved accuracy by 6.8 percent.

Graduate student Jiang Xue is the paper's first author. The study was supported by the National Key R&D Program.

Structural diagram of the PlantCaFo model. [Photo/news.hzau.edu.cn]