HCA-CLIP: Hierarchical context-aware adaptation for few-shot plant disease recognition

Few-shot plant disease recognition remains challenging because different diseases may exhibit highly similar visual symptoms, while the appearance of the same disease can vary considerably under different imaging conditions. Although CLIP provides transferable vision-language representations, its generic pre-training tends to emphasize high-level semantics while overlooking the fine-grained pathological cues required for distinguishing visually similar diseases. To address this limitation, we propose HCA-CLIP, a hierarchical context-aware adaptation framework that progressively bridges the gap between generic vision-language knowledge and disease-specific visual representations. HCA-CLIP first recovers complementary fine-grained information from intermediate visual features through hierarchical differential feature refinement (HDFR). Multi-view supervised contrastive learning (MVSCL) is then employed to improve the compactness and separability of the refined representations, while adaptive de-confusion (ADC) further calibrates predictions by dynamically coordinating generic CLIP knowledge with disease-specific visual evidence. Extensive experiments on the PlantVillage, AI Challenger 2018, and SPD datasets demonstrate consistent improvements across different few-shot settings. Under the challenging 1-shot setting, HCA-CLIP achieves accuracies of 69.58 ± 0.35%, 54.05 ± 0.39%, and 44.84 ± 0.42%, outperforming the strongest evaluated competitors by 6.26, 1.20, and 3.10 %age points, respectively. These results demonstrate the effectiveness of progressively refining visual representations and calibrating prediction boundaries for fine-grained plant disease recognition under limited supervision.