A lightweight graph-enhanced deep learning framework for explainable cucumber leaf disease diagnosis

Accurate and efficient identification of cucumber leaf diseases is a critical step in preventing losses and facilitating timely intervention in agricultural activi-ties. However, most state-of-the-art plant disease recognition models, including those employing deep learning, often fail to identify spatial dependencies among symptomatic leaf feature regions, require high computational resources, and lack robustness in their predictions. To overcome these challenges, this paper pro-poses MobileGraph, a graph-aided deep learning model that jointly reasons local texture patterns and spatial dependencies among CNN-derived cucumber leaf feature regions using MobileNetV3 as a lightweight feature extractor. Experi-ments on a publicly available cucumber leaf disease dataset containing 5 classes and 4,000 images show that the proposed model achieves 99.75% accuracy, 99.75% macro F1-score, and 99.69% MCC, outperforming several state-of-the-art models including ResNet-152, EfficientNet-B7, DenseNet-201, ConvNeXt, and VGG16, while having a significantly lower computational cost of 0.465 GFLOPs. Explainability results from Grad-CAM and LIME indicate that the model is focused on biologically important regions of plant lesions. Furthermore, a proto-type mobile application illustrates the feasibility of real-time cucumber disease diagnosis for practical agricultural monitoring. These results indicate that Mobi-leGraph provides an efficient and interpretable solution for intelligent crop health surveillance.