Plant diseases pose a significant threat to global food security, resulting in annual yield losses of 20–40% worldwide. Early and accurate detection is essential for advancing precision agriculture and promoting sustainable crop management. This PRISMA-guided systematic review synthesizes findings from 155 peer-reviewed studies (2020-2025) examining advances in Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and hybrid CNN-ViT architectures for image-based plant disease detection. The review evaluates data acquisition strategies, preprocessing pipelines, and the integration of hyperspectral and multimodal imaging modalities. Findings indicate that CNNs remain strong baselines, achieving 90-99% accuracy in controlled settings (as reported in reviewed studies). In comparison, hybrid CNN-ViT architectures demonstrate superior robustness (>97%) and generalization in heterogeneous field environments. Persistent challenges include dataset imbalance, limited annotated hyperspectral data, high computational demand, and constrained cross-domain adaptability. Future frameworks should prioritize lightweight, edge-deployable, and explainable AI models supported by standardized multimodal datasets. Overall, this review consolidates recent progress, highlights critical research gaps, and proposes a roadmap toward interpretable, scalable, and energy-efficient deep learning solutions that can enable real-time, field-ready disease detection for sustainable agriculture.