Knowledge distillation for smart agriculture: methods, applications, and future directions

Knowledge Distillation (KD) has emerged as a powerful paradigm for optimizing deep learning models in agriculture, enabling the transfer of knowledge from complex teacher networks to lightweight student models. Following PRISMA guidelines, this review surveys 92 studies across key domains, including plant disease recognition, crop and weed detection, pest recognition, and animal monitoring. To address the lack of standardized benchmarking across heterogeneous agricultural tasks, we adopt the Performance Retention Ratio (PRR) as a novel metric for quantifying distillation efficiency. We evaluate the efficacy of response-based, feature-based, and relation-based KD strategies, identifying critical challenges, including data scarcity, environmental variability, mismatched teacher and student architectures, limited interpretability, and dynamic agricultural conditions. The review further highlights emerging solutions, such as adaptive distillation, hardware-aware distillation, cross-domain learning, explainable distillation, dynamic distillation and lifelong learning frameworks. Our findings underscore the growing potential of KD to advance precision and sustainable agriculture by enabling efficient, interpretable, and field-deployable artificial intelligence systems.