MSPRP: multi-granularity symptom-aware probabilistic representation prompting for target-domain-free plant disease detection

Introduction Existing vision-language zero-shot anomaly detection methods have mainly been validated on industrial datasets and often suffer from lesion-scale variations, domain shifts, and complex background interference when directly applied to field plant images. To address these challenges, this paper proposes MSPRP, a multi-granularity symptom-aware probabilistic representation prompting method for target-domain-free plant disease detection. Methods Based on frozen CLIP encoders, MSPRP introduces three modules: the Multi-Granularity Symptom-Aware Module (MGSAM) performs learnable hierarchical fusion and spatial-channel symptom attention on multi-layer patch features, the Uncertainty-Aware Prompt Flow Module (UPFM) models healthy/diseased prompts as probabilistic representations to improve prompt adaptability, and the Adaptive Cross-Modal Prototype Aggregation Module (ACPAM) aggregates suspected symptom prototypes from high-response patches to reduce background interference in image-level disease scoring. The model is trained only on the industrial MVTec AD dataset and tested on target agricultural datasets without target-domain training. Results A field Polygonatum odoratum disease dataset is also constructed as an evaluation benchmark. Experimental results show that MSPRP achieves the best average AUROC and AP among methods that do not use target agricultural training images. Ablation results verify the contribution of each module. Discussion Overall, MSPRP provides a target-domain-free computational methodology for field plant disease discrimination without target agricultural training images or target-domain pixel-level lesion annotations.