AdjLeafGNN: a hybrid deep learning and graph neural network framework for probabilistic modeling of adjacent leaf disease spread in precision agriculture

Proper detection and treatment of plant leaf diseases are essential factors for achieving good crop yields and ensuring food security. Convolutional Neural Networks (CNNs) have shown significant potential for classifying diseases from leaf images. Instead, most current work focuses on image-level prediction and ignores the relationship between infected leaves. This limitation somewhat constrains their use in modelling disease spread. Also, it makes them less efficient in typical field situations where disease is transmitted from plant to plant by physical contact. Moreover, existing CNN architectures do not access inter-lobar contextual information, an essential factor for early detection and control. To tackle this, we propose AdjLeafGNN, an innovative hybrid deep learning and graph neural network model that performs multi-class leaf disease classification and probabilistic prediction of adjacent-leaf disease spread in a single pass. The method uses the enhanced CNN model (LDDNet), with Atrous Spatial Pyramid Pooling (ASPP) and a Channel-Spatial Attention Module (CSAM), to achieve a more precise representation across multiple scales. These embeddings are then used to construct a similarity graph, enabling a GNN to infer likely disease transmission paths among leaves. We evaluate the PlantVillage dataset on the proposed model, and the results show that it outperforms state-of-the-art CNN-based methods, achieving 98.88% classification accuracy and 98.71% F1 Score. Additionally, we were able to predict disease spread with a high AUC-ROC of 0.942 and an MCC of 0.884 using our framework. These results confirm that AdjLeafGNN can accurately model both local and relational patterns. The approach we propose is scalable and interpretable, facilitating real-time monitoring and control of diseases in precision agriculture.