An interpretable generative feature fusion network with hybrid optimization for plant disease classification: GGMHO.Net

Plant diseases continue to severely impact global agriculture by causing extensive crop losses. AI and IoT-based smart farming technologies are being adopted to facilitate early disease detection. However, the limited generalizability of existing deep learning models, their unreliable decision-making processes, and the inadequacy of robust in-field datasets present critical challenges to their effective and scalable implementation. This study proposes a GGMHO.Net (Generative-GCN Multi-scale Hybrid Optimization Network) model for plant disease classification, with enhanced generalizability and explainability. The architecture of the proposed model integrates Graph Convolutional Network (GCN) and Multi-Scale Convolution (MSC) networks to effectively capture both local and global features. Furthermore, it incorporates a Wasserstein Distance Metric (WDM) for refined feature extraction and employs a novel hybrid optimization algorithm, SSA-PSO, for feature selection. To support this approach, we introduce the Lettuce Leaf Disease Dataset (LLDS) consisting of four classes obtained from our experimental set-up. Proposed model attains a true positive rate (TPR) of 0.9981 and F1-score of 0.9981 on LLDS, outperforming other state-of-the-art crop disease detection models. Notably, the system is validated on benchmark datasets, yielding strong results: TPRs of 0.9970 (lettuce), 0.9957 (lettuce NPK), and 0.9920 (potato). Additionally, the model provides explainable predictions via SHAP. With GFLOPs of 5.44, and frames per second (FPS) of 29.07, the proposed model can be efficiently deployed in resource-constrained agriculture systems.