Deep learning and GCC-MA based visual system for mango leaf disease analysis with anthracnose and sooty mold as case study

  Full text PDF

Mango leaf diseases threaten yield and quality, underscoring the need for early diagnosis and actionable monitoring. Addressing the absence of an integrated Visual Management System (VMS) in prevailing methodologies, this work introduces an enhanced VMS for real-time mango leaf disease analysis that unifies computer vision and language understanding. The framework employs Soft Root Sign–VGG16–Fast Dropout–ResNet50 (SRSVGG16FDResNet50) for visual classification and Generalized Canonical Correlation Multimodal Autoencoders (GCCMA) for cross-modal feature fusion. On the visual side, mango leaf images were preprocessed, followed by leaf color detection, superpixel segmentation, and feature extraction. In parallel, disease-related question–answer (Q&A) data were curated and preprocessed; keywords are extracted, entity–relation graphs were constructed, and textual features were derived. LeCun Bayesian BERT (LeCunBayBERT) generates embeddings for answers, questions, and keywords to enhance semantic representation. GCCMA was used to fuse the visual and textual to model crossmodal correlations and finally, the disease classification was done by SRSVGG16FDResNet50. With testing, the user is authenticated and provides images with queries via the VMS interface, and the system provides the predicted diagnosis as well as the natural language responses. The proposed framework outperforms existing methodologies for mango leaf disease detection, achieving higher accuracy (98.45%) while supporting interpretable, real-time decision support for orchard management. As a case study, we demonstrate the system’s effectiveness on anthracnose and sooty mold, illustrating end-to-end detection and interpretation.