LeafVRNet: An Explainable Feature Fusion Framework for Interpretable Plant Leaf Disease Detection

Abstract

Early detection of plant leaf diseases is a key factor in sustainable crop management and minimizing agricultural losses. Accurate and timely identification is particularly important for ensuring food security in regions with limited agricultural expertise, such as Bangladesh. This paper introduces LeafVRNet, a hybrid Deep Learning model combining VGG16 and ResNet50V2, for multi-crop leaf disease classification. The dataset consists of 21 disease categories from six species, collected from local agricultural fields in Bangladesh. To enhance image quality, preprocessing techniques such as Background Removal, Region of Interest (ROI) extraction and CLAHE contrast enhancement have been applied. Multiple pre-trained Convolutional Neural Network (CNN) models including VGG16, ResNet50V2, VGG19, MobileNetV2, InceptionV3 and Xception have been assessed, where VGG16 has achieved the highest accuracy of 97.35 %, followed closely by ResNet50V2 with 97.27 %. LeafVRNet has achieved the best performance among tested models with 98.83% accuracy, alongside high precision, recall and F1-score. Local Interpretable Model-Agnostic Explanations (LIME) technique has been employed to highlight the regions of leaf images that influence predictions the most. The results confirm that combining accurate DL models with XAI can create a reliable and transparent disease detection system. This approach supports the early diagnosis and prevention of crop diseases, contributes to sustainable agriculture, reduces economic losses and improves food security.

Keywords