Hybrid deep learning for smart paddy disease diagnosis using self supervised hierarchical reconstruction and attention based temporal analysis

Accurate and early disease detection in paddy crops is essential for maximizing crop yield which ensures food security. Traditional methods are often labor-intensive, time-consuming, and domain-specific expertise. Feed-forward deep-learning models will perform accurate disease detection through the identification of spatial patterns. However, they cannot predict the diseases at the early stages due to the lack of temporal information. Temporal observations will help perform continuous monitoring and detect minute changes in the crops at the early times. To tackle this problem, we proposed Self-Supervised Deep Hierarchical Reconstruction (SSDHR), and Long Short-Term Memory (LSTM) which perform early disease detection based on the spatial and temporal data respectively. The SSDHR network uses multi-branch convolution kernels to extract distinct discriminative characteristics rather than conventional leaf-based indicators. It incorporates spatial, and temporal-based attention mechanism Symmetric Fusion Attention (SFA) to improve feature selection and XGBoost (XGB) classifier for better stability. According to experimental findings, the suggested framework achieves a 99.25% accuracy rate in identifying and classifying 13 paddy classes, including normal, blast, hispa, tungro, white stem borer, brown spot, leaf roller, downy mildew, yellow stem borer, bacterial leaf blight, bacterial leaf streak, black stem borer, and bacterial panicle blight.