Abstract
Plant disease image classification is a key task in disease monitoring and precise prevention and control in smart agriculture. However, complex backgrounds, finegrained lesion differences, and model overconfidence still limit recognition performance and practical application reliability. To address these issues, this paper proposes a reliable plant disease recognition framework that integrates multi-frequency selective state-space feature enhancement with conformal-aware reliable prediction. The proposed method adopts Swin-Tiny as the backbone network and uses the MF-SSFE module to jointly model low-frequency leaf structures, high-frequency lesion textures, and cross-region state-space contexts, thereby enhancing disease-related evidence in complex scenarios. Meanwhile, the CARP module is introduced to incorporate the idea of conformal prediction into the classification output process, enabling the model to express uncertainty while providing class predictions. Experimental results show that the proposed method achieves Accuracy values of 0.9891, 0.3889, and 0.9800 on the NGLD, PlantDoc, and PlantVillage datasets, respectively, and AUC values of 0.9995, 0.8690, and 0.9989, respectively, outperforming the comparison methods. Ablation experiments and sensitivity analysis further verify the effectiveness and stability of each module. Model complexity analysis shows that the proposed method maintains acceptable computational overhead while achieving superior recognition performance.