Diagnosis of banana plant disease is a crucial aspect of sustaining the harvest of crops and their quality. Visual inspection of certain diseases like Black Sigatoka, Panama disease, and aphids is not easy and can lead to misjudgments. Generally, traditional deep learning approaches have been previously used but they have not performed well in addressing issues of class imbalance, sensitive disease differentiation and noisy images obtained in the field. Furthermore, most models are based on a collection of predetermined preprocessing methods and single-path networks that limit their ability to generalize to a wide variety of environments. Current methods of deep learning tend to achieve reasonable overall performance but fail to perform well on key performance indicators such as recall and F1-score when considering underrepresented and overlapping classes, such as Yellow and Black Sigatoka. Such constraints impede efficient field implementation, as diseases of minority classes are often falsely classified. To overcome these deficiencies, we develop a novel Duel-Path Attention Fusion Network (DPAFNet) that is trained utilizing adaptive quantum monarch butterfly optimization (AQMBO). The concept behind the proposed model is to feed MaxViT and HorNet-S two feature extractors to deliver global contextual details and minute-scale textural features. The traditional filters which do a reasonable job in handling dynamic noise and contrast are replaced by a learnable preprocessing unit. The cross-layer fusion attention encourages interclass discriminative learning of diseased plants. The suggested model has been trained and tested on an open-source dataset of Mendeley banana disease, which includes 5,170 images in 7 disease categories and 1 control condition. The accuracy, F1-score and MCC of 98.6% and 0.93 and 0.87 respectively (achieved experimentally) demonstrate the superiority of DPAFNet over baseline models such as EfficientNetB0 (accuracy 95.0%), DenseNet121 and ResNet50 (accuracy 93.50% and 92.0% respectively). As can be seen, the model had a 0.26-0.48 increase in F1-score in the challenging Panama disease category. These results prove that the proposed architecture can be successfully used to achieve high-accuracy disease classification in smart agriculture that is robust and prepared for field implementation.

Dual-Path Attention Fusion Network (DPAFNet) Adaptive Quantum Monarch Butterfly Optimization (AQMBO) MaxViT HorNet-S Cross-Layer Attention Fusion (CLAF) Banana Leaf Disease Classification