Citrus leaf pathology multi-class image dataset and a hybrid multi-branch attention fusion network for lemon leaf disease classification

Abstract The productivity of lemon (Citrus lemon) crops is severely undermined by foliar diseases, which constitute the leading cause of global harvest degradation. The conventional reliance on manual identification by experienced specialists poses a significant bottleneck, as these methods are both time-consuming and prone to human error under field conditions. Recently, deep learning techniques have been employed to identify and diagnose citrus foliar diseases from digital images, automating the classification process and providing non-experts with accessible diagnostic tools. While several deep learning applications have been used to classify lemon diseases with the aim of improving classification accuracy, most of these studies fail to achieve high accuracy and rely on datasets with an insufficient number of disease classes. In this paper, we introduced a dataset named “Citrus Leaf Pathology Multi-Class Image Dataset” with 14 different classes, and we proposed a Hybrid Multi-Branch Attention Fusion Network (HMAF-Net). The architecture consists of four parallel feature extraction branches, each incorporating branch attention and channel attention, followed by classification using a Neural Network (NN). A DenseNet121 backbone is combined with lightweight convolutional branches, refined through squeeze-and-excitation channel attention and a learnable branch-weighting fusion mechanism to adaptively integrate complementary feature representations. The model attains 96.63% precision and 95.59% accuracy on the proposed dataset. The model further demonstrates strong generalization, achieving 97.48%, 95.89%, and 98.23% accuracy on three independent external lemon and citrus leaf datasets, with statistical significance confirmed through paired t-tests across five independent runs. Explainable AI visualizations further confirm that the model’s predictions are grounded in disease-relevant leaf regions, supporting its potential for reliable deployment in real-world agricultural diagnostic systems.