HAFNet: A Hierarchical Attention-Based Deep Learning Model for Mango Leaf Disease Detection

This study introduces the Mango Branch - Multi Attention Fusion Network (MAFN) framework for the identification of mango leaf diseases, employing feature engineering techniques such as color histograms, texture analysis, and shape descriptors. The research implemented data preprocessing strategies and utilized a multi-branch architecture with an attention mechanism to extract hierarchical features from images. The model demonstrated robust performance, achieving 92% accuracy, 94% precision, 93% recall, and 94% F1 score. Additionally, a second model, the Hierarchical Attention Based Fusion Network (HAFNet), was developed to integrate multiple branches for multi-scale feature extraction while mitigating noise through attention-based features. The study advanced preprocessing methods for both MAFN and HAFNet, introducing novel attention-based features in the field of agricultural science. The findings indicate enhanced crop disease detection through feature engineering, thereby promoting sustainable farming management via early detection. The paper concludes by discussing the limitations of the models and potential avenues for future research.