Detection of citrus diseases in complex backgrounds based on image-text multimodal fusion and knowledge assistance

Diseases pose a significant threat to the citrus industry, and the accurate detection of these diseases represent key factors for their early diagnosis and precise control. Existing diagnostic methods primarily rely on image models trained on vast datasets and limited their applicability due to singular backgrounds. To devise a more accurate, robust, and versatile model for citrus disease classification, this study focused on data diversity, knowledge assistance, and modal fusion. Leaves from healthy plants and plants infected with 10 prevalent diseases (citrus greening, citrus canker, anthracnose, scab, greasy spot, melanose, sooty mold, nitrogen deficiency, magnesium deficiency, and iron deficiency) were used as materials. Initially, three datasets with white, natural, and mixed backgrounds were constructed to analyze their effects on the training accuracy, test generalization ability, and classification balance. This diversification of data significantly improved the model's adaptability to natural settings. Subsequently, by leveraging agricultural domain knowledge, a structured citrus disease features glossary was developed to enhance the efficiency of data preparation and the credibility of identification results. To address the underutilization of multimodal data in existing models, this study explored semantic embedding methods for disease images and structured descriptive texts. Convolutional networks with different depths (VGG16, ResNet50, MobileNetV2, and ShuffleNetV2) were used to extract the visual features of leaves. Concurrently, TextCNN and fastText were used to extract textual features and semantic relationships. By integrating the complementary nature of the image and text information, a joint learning model for citrus disease features was achieved. ShuffleNetV2 + TextCNN, the optimal multimodal model, achieved a classification accuracy of 98.33% on the mixed dataset, which represented improvements of 9.78% and 21.11% over the single-image and single-text models, respectively. This model also exhibited faster convergence, superior classification balance, and enhanced generalization capability, compared with the other methods. The image-text multimodal feature fusion network proposed in this study, which integrates text and image features with domain knowledge, can identify and classify citrus diseases in scenarios with limited samples and multiple background noise. The proposed model provides a more reliable decision-making basis for the precise application of biological and chemical control strategies for citrus production.