Plant leaf disease classification and severity grading are essential for precision agriculture, enabling timely intervention and optimized management. Existing models often fail to recognize previously unseen disease categories due to rigid label spaces and limited representation of plant phenotypes. To address these challenges, we propose a knowledge-driven unified framework for plant disease classification and severity grading. A Common Knowledge Learner consolidates fundamental features of plant species, disease categories, and severity levels from labeled data, forming a transferable representation space. It employs a multi-level contrastive learning strategy to capture both global semantic representations and fine-grained lesion patterns. Building on these representations, a Cross-Domain Adaptation module leverages a teacher–student framework with Low-Rank Adaptation (LoRA) bridges in-domain and out-of-domain feature spaces using large-scale unlabeled data. Meanwhile, a contrastive feature library enables similarity-based reasoning and supports flexible label space expansion during inference without retraining. We evaluate our approach on Leaf-CG, a large-scale dataset comprising 441,448 images from 59 plant species, 373 disease categories, and four severity levels. Experiments demonstrate that our framework outperforms existing baselines, achieving 94.9% disease classification accuracy and 90.6% severity grading accuracy in-domain. Under out-of-domain conditions, the method achieves 82.1% true positive rate (TPR) in open-set settings, highlighting its strong generalization ability and potential for practical plant disease management. Code and dataset are available at https://www.uniplantcg.samlab.cn.

Plant diseases remain a major constraint on crop productivity, requiring timely and accurate diagnostic approaches to secure agricultural yields. While existing automated diagnosis methods primarily rely on image data and achieve notable results, their performance often declines in complex field environments with noise and interference. Multimodal learning provides a promising solution by integrating complementary cues from various data sources. However, the heterogeneity between plant phenotypes and other modalities, such as textual descriptions, poses a significant challenge for effective fusion. To address this issue, we propose PlantIF, a multimodal feature interactive fusion model for plant disease diagnosis based on graph learning. PlantIF comprises three key components: image and text feature extractors, semantic space encoders, and a multimodal feature fusion module. Specifically, we employ pre-trained image and text feature extractors to extract visual and textual features enriched with prior knowledge of plant diseases. Semantic space encoders then map these features into both shared and modality-specific spaces, enabling the capture of cross-modal and unique semantic information. To enhance context understanding, we design a multimodal feature fusion module to process and fuse different modal semantic information, and then extract the spatial dependency between plant phenotype and text semantics through the self-attention graph convolution network. We evaluate PlantIF on a multimodal plant disease dataset with 205,007 images and 410,014 texts, achieving 96.95 % accuracy, 1.49 % higher than existing models. These results demonstrate the potential of multimodal learning in plant disease diagnosis and highlight PlantIF's value in precision agriculture. Codes are available at https://github.com/GZU-SAMLab/PlantIF.