EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

CA-CLIP for multi-label diagnosis of co-occurring plant leaf diseases: A class-aware vision-language framework

Shuchen Liu, Jinming Wu, Wenqiang Zhao, Juan Liu, Jing Chen, Yunpeng Cui

Smart Agricultural Technology · 2026-12 · 15 · 102617

DOI: 10.1016/j.atech.2026.102617 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Smart Agricultural Technology, 2026-12. Verfasst von Shuchen Liu und 5 weiteren Autor:innen. Die Arbeit nennt 59 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: National Key Research and Development Program of China, Chinese Academy of Agricultural Sciences.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate recognition of co-occurring plant leaf diseases is important for crop health monitoring in agriculture. However, most existing image-based disease diagnosis models are designed for single-disease classification and are less suitable for leaves exhibiting mixed symptoms. To address this limitation, this study proposes CA-CLIP, a class-aware vision-language framework for image-level multi-label diagnosis of symptom-mixed leaves. A controlled co-occurring plant leaf disease (CCPLD) benchmark was constructed from five crop species, comprising 3,400 synthetic co-occurring disease images covering 17 predefined co-occurring disease types, together with single-disease and healthy leaf images for controlled evaluation. To introduce image-verifiable disease semantics, a structured prompt bank was developed from visible leaf symptom descriptions. CA-CLIP integrates class-specific text prototypes with visual features through prompt-guided feature modulation and class-aware gating (CA-gating), enabling independent prediction of multiple disease labels from the same image. On the CCPLD validation set, the visible-symptom prompt configuration achieved the best ablation performance, with subset accuracy of 0.8543 and Macro-F1 of 0.9601. CA-CLIP outperformed representative image-only and vision-language baselines, improving subset accuracy by 4.39 percentage points over CLIP-Adapter. For lightweight inference, logit-based knowledge distillation transferred the diagnostic capability of CA-CLIP to an image-only MobileNetV3 student, which achieved subset accuracy of 0.8066 with 2.99 M parameters. Under the single-label evaluation protocol, the distilled model achieved a pooled target-label recognition rate of 0.5156 and a pooled mAP of 0.5626 across eleven disease categories in the two natural-background external evaluation sets, indicating a degree of cross-domain transferability from the controlled CCPLD benchmark to natural-background images. This study provides a controlled benchmark and a transparent proof-of-concept framework for symptom-mixed plant disease diagnosis, with potential for lightweight local inference.

Förderung

  • National Key Research and Development Program of China
  • Chinese Academy of Agricultural Sciences

Zitieren

APA

Liu, S., Wu, J., Zhao, W., Liu, J., Chen, J., & Cui, Y. (2026). CA-CLIP for multi-label diagnosis of co-occurring plant leaf diseases: A class-aware vision-language framework. Smart Agricultural Technology, 15, 102617. https://doi.org/10.1016/j.atech.2026.102617

Vancouver

Liu S, Wu J, Zhao W, Liu J, Chen J, Cui Y. CA-CLIP for multi-label diagnosis of co-occurring plant leaf diseases: A class-aware vision-language framework. Smart Agricultural Technology. 2026;15:102617. doi:10.1016/j.atech.2026.102617

BibTeX
@article{Liu2026,
  author = {Liu, Shuchen and Wu, Jinming and Zhao, Wenqiang and Liu, Juan and Chen, Jing and Cui, Yunpeng},
  title = {CA-CLIP for multi-label diagnosis of co-occurring plant leaf diseases: A class-aware vision-language framework},
  journal = {Smart Agricultural Technology},
  year = {2026},
  volume = {15},
  pages = {102617},
  doi = {10.1016/j.atech.2026.102617}
}
Kostenlos starten →