LeafFusionNet: a hybrid deep learning approach for robust plant disease detection

Introduction: Crop diseases have to be diagnosed early to save crop yields and food safety. Commonly used conventional practices are time-consuming and likely to involve human error. Automatic plant disease detection is a highly efficient technology that has been brought into existence by the development of deep learning and computer vision technology and can effectively detect the signs of diseases in plant species. Sustainable agriculture and early intervention depend on accurate and interpretable detection of plant diseases. Methods: This study introduces a hybrid model based on deep learning techniques that effectively identifies and categorizes leaf diseases. The proposed model, LeafFusionNet, incorporates Convolutional Neural Network (CNN) and Vision Transformer (ViT) with an efficient attention module, LeafTAM (Leaf Texture Attention Module), to effectively capture both global and local information. This architecture is enhanced by the addition of a Gabor filter layer before the CNN-ViT fusion, hence augmenting the model's ability to extract physiologically relevant texture features. Results and discussion: The model was trained and validated on the Plant Village dataset. This model attained an accuracy of 99.33%, 99% precision, recall, and F1 score, demonstrating strong generalization on new data when compared with the state-of-the-art models. This proposed hybrid model can be utilized to develop a strong agricultural diagnostics system. These results point to the possibility of designing a powerful, interpretable system by utilizing transformer-based vision modules, Gabor filters, and explainability systems such as Grad-CAM.