Explainable and Lightweight Deep Learning Models for Tea Leaf Disease Detection: A Systematic Review and Future Research Directions

• Systematically reviews 76 studies on tea leaf disease detection using deep learning • Introduces a unified taxonomy for lightweight and explainable model design • Compares model accuracy, latency, and efficiency across major architectures • Identifies gaps in dataset quality, generalization, and explainability fidelity • Outlines future directions for deployable, interpretable, and mobile-friendly systems Tea is the second most widely consumed beverage in the world, and its production and quality are economically significant. Therefore, it is essential to monitor tea leaf health to protect the yield; however, manual inspection is labor-intensive and often inefficient. Recent advances in deep learning have improved the accuracy and feasibility of automated tea-leaf disease identification. Regardless of this development, significant challenges remain, particularly regarding in-field applicability and compatibility with devices with limited resources. This systematic review synthesizes deep learning approaches for tea-leaf disease detection, covering studies from 2015 to 2025. Following PRISMA guidelines, an initial search of five databases yielded 524 records; 76 peer-reviewed studies were included after screening. The review categorizes key plant stresses, tabulates public datasets, compares deep learning architectures, and evaluates lightweight models, XAI, and real-world variability. A methodological quality assessment using a predetermined checklist confirms that the literature’s quality has advanced. The literature shows notable achievements in accuracy and clear progress toward on-device inference. However, the field is still hampered by a lack of ideal datasets, limited model explainability, poor deployment reporting, and insufficient standardized evaluation for efficiency and robustness. As a solution to these gaps, this review proposes an evidence-based set of design guidelines for lightweight and explainable recognition of tea leaf diseases, emphasizing external validation, transparent explanations, and deployment-ready efficiency reporting; empirical benchmarking under a unified protocol is identified as a priority for future work. This review on recognizing tea leaf diseases aims to assist researchers, support farmers, identify gaps, and provide recommendations for future developments in precision tea agriculture.