Agraph neural network framework for vascular streak dieback recognition

Abstract

Vascular streak dieback (VSD) is one of the most destructive diseases affecting cocoa production in Southeast Asia, including Indonesia, where early visual symptoms are often subtle and spatially distributed across the leaf sur face. Conventional image-based disease recognition approaches, particularly those relying solely on convolutional neural networks (CNNs), are effective in extracting local visual features but remain limited in modeling long-range structural relationships such as venation disruption and lesion spread. To ad dress this limitation, this study investigates a hybrid CNN-graph neural network (CNN-GNN) framework for automated VSD recognition from cocoa leaf im ages. A primary dataset consisting of 1,000 RGB images collected directly from cocoa plantations in Jember Regency was used to reflect realistic field condi tions. In the proposed approach, CNNs are employedfor local feature extraction, while graph-based representations enable GNNs to capture global relational pat terns through message passing. Experimental results demonstrate stable learning behavior and strong classification performance, achieving a maximum validation accuracy of 95.2% and an area under the curve (AUC) of approximately 0.94. Further analysis shows balanced precision and recall across classes, indicating reliable discrimination between Sehat and VSD-infected leaves. These findings suggest that hybrid CNN-GNN modeling provides an effective strategy for cap turing both local and distributed structural characteristics of VSD symptoms and highlights the potential of graph-based reasoning to complement convolutional feature learning in plant disease diagnostics.