Early detection of crop diseases is critical for safeguarding yields, but collecting large, high-quality field datasets is expensive, time-consuming, and often impossible for rare plant diseases.
A new study published in Plant Phenomics (DOI: 10.1016/j.plaphe.2025.100024) by Ruifang Zhai’s team at Huazhong Agricultural University introduces PlantCaFo, a few-shot learning model designed to accurately identify plant diseases even with limited samples.
By integrating a lightweight Dilated Contextual Adapter (DCon-Adapter) and a Weight Decomposition Matrix (WDM), PlantCaFo achieves 93.53% accuracy in controlled tests and surpasses existing approaches in real-world scenarios. The model leverages pretrained foundation networks — CLIP, DINO, and DINO2 — while training only essential components for efficiency, according to a press release.
Experiments on benchmark datasets, including PlantVillage and Cassava, show consistent improvements of up to 7.74% in accuracy. Visual analyses confirm that PlantCaFo effectively focuses on disease-affected regions while minimizing distractions from background features.
This technology could transform crop protection by enabling rapid, low-cost disease diagnostics. Farmers, agronomists, and plant health agencies can deploy AI-based detection tools on mobile apps, drones, or early-warning systems to curb outbreaks, reduce crop losses, and enhance global food security.
