Attention-enhanced hybrid deep learning model for robust mango leaf disease classification via ConvNeXt...

Mango is a crop of vital agronomic and commercial importance, particularly in tropical and subtropical regions. Accurate and timely identification of foliar diseases is essential for maintaining plant health and ensuring sustainable agricultural productivity. This study proposes MangoLeafCMDF-FAMNet (cross-modal dynamic fusion with feature attention module (FAM) network), an advanced, hybrid, deep-learning framework designed for the multi-class classification of mango leaf diseases. The model combines two state-of-the-art feature extractors, ConvNeXt and Vision Transformer, to capture local fine-grained textures and global contextual semantics simultaneously. To further improve feature discrimination, a FAM inspired by squeeze-and-excitation networks is integrated into each stage of the backbone. This module adaptively recalibrates channel-wise feature responses to highlight disease-relevant cues while suppressing irrelevant background noise. A novel cross-modal dynamic fusion strategy unifies the complementary strengths of both branches, resulting in highly robust and discriminative feature embeddings. The proposed model was rigorously evaluated using comprehensive metrics such as classification accuracy (CA), recall, precision, Matthews correlation coefficient (MCC) and Cohen's kappa score on three benchmark datasets: MangoLeafDataset1 (8 classes), MangoLeafDataset2 (5 classes) and MangoLeafDataset3 (8 classes). The experimental results consistently demonstrate the superiority of MangoLeafCMDF-FAMNet over the existing baseline models. It achieves exceptional CA values of 0.9978, 0.9988 and 0.9943 across the respective datasets, alongside strong MCC and Cohen's kappa scores. These results highlight the effectiveness and generalizability of the proposed framework for automated mango leaf disease diagnosis and contribute to advancing deep learning applications in precision plant pathology.