Lightweight Deep Learning for Real-Time Plant Disease Detection on Mobile and Edge Devices: A Comprehensive Review and the TinyCNN-Lite Framework

Plant diseases cause 20–40% of annual crop losses globally, threatening food security for millions of smallholder farmers in developing regions. Recent advances in lightweight deep learning and mobile edge computing present a compelling opportunity to deploy expert-level plant disease diagnosis on sub-$100 smartphones. This review systematically surveys the landscape of deep learning approaches for plant disease detection, encompassing CNN architectures, attention mechanisms, model compression techniques (pruning, knowledge distillation, quantization), neural architecture search, explainable AI, and federated learning. A five-dimensional taxonomy organizes 25+ reviewed works across architecture design, compression, training strategy, edge deployment, and interpretability. Building on this synthesis, we present TinyCNN-Lite, a novel lightweight framework integrating adaptive bottleneck inverted residuals, efficient hybrid Attention with less than 2% overhead, a four-stage compression pipeline achieving greater than 10× size reduction, decentralized federated learning with differential privacy, and Grad-CAM farmer-centric explainability. A comprehensive one-year research plan with monthly milestones, a three-phase structure, and concrete deliverables is provided. Critical research gaps in domain generalization, data scarcity, energy efficiency, and continual learning are identified. Targets include a model below 5 MB, inference under 50 ms on MediaTek MT6765, and accuracy exceeding 95% across 50 disease classes from 10 major crops.