Accurate and timely plant disease identification is critical for global food security, yet automated methods often face challenges in effectively integrating diverse feature types. This paper introduces a hybrid deep learning architecture for precision agriculture, combining visual features from advanced CNNs such as ConvNeXtV2-L, Swin-L, and ResNet152d with handcrafted structural features. Feature integration is achieved through a sophisticated multi-stage process involving attention-based adapters, a transformer encoder for inter-backbone fusion, and a multi-head attention mechanism for final hybrid modality fusion. Evaluated on the standard 38-class Plant Village dataset test split, the proposed model demonstrates exceptional performance, achieving 99.85% accuracy and a 99.85% weighted F1-score. This work highlights the efficacy of combining diverse feature sources through advanced, attention-guided fusion techniques, presenting a highly accurate and robust solution for automated plant disease diagnosis in precision agriculture.