A Dynamic Attention-Guided Deep Learning Framework (D-PAN) With Reduced Complexity for Fine Grained Severity-Aware Classification of Micronutrient Deficiency and Disease Detection in Banana and Coffee Leaves

Nutrition is a key factor in enabling crops to reach their optimal growth level. Farmers and agronomists often mistake nutrient deficiencies for diseases, leading to incorrect crop treatments. This study proposes a deep learning model for classifying micronutrient deficiencies (m-ND), evaluating their severity, and distinguishing them from diseases in banana and coffee crops. To address common classification challenges, the model is initially trained and validated on images of both nutritional deficiencies and visually similar disease symptoms. To assess its robustness, we also evaluated it on a coffee crop dataset from Mendeley. The proposed method, D-PAN - a modified Pyramid Attention Network (PAN) uses a dynamic attention mechanism to process multi-scale input images. Model performance was compared with various vision transformer (ViT) based channel attention methods and a weighted ensemble of Transformer models combined with a modified convolution neural network (CNN). The model achieved top accuracies of 98.80% for banana and 94.60% for coffee, with reduced misclassification errors. Experimental results show that the framework maintains low complexity, improves accuracy by an average of 2.9%, and reduces parameters by 1.64 to 3.83 times compared to other attention methods. Overall, the proposed model surpasses all evaluated state-of-the-art attention mechanisms, achieving strong performance across recall, accuracy, precision, and F1-score.