抄録
The rapid expansion of artificial intelligence (AI) in crop growth analysis has produced fragmented advances across multimodal sensing, phenotypic perception, spatiotemporal prediction, and decision support. The central challenge is no longer the availability of AI applications, but understanding how heterogeneous observations can be transformed into reliable crop-state information and agronomically meaningful decisions, and under what conditions such intelligence remains transferable across crops, environments, and management regimes. This review addresses this problem through an integrated framework that traces the evolution of crop growth AI from data and perception to prediction, decision, and agricultural value, linking sensing modalities, model architectures, biological processes, and deployment contexts. Comparative synthesis shows that increasing model complexity does not necessarily improve agricultural value, because predictive performance is constrained by data quality, environmental representativeness, crop–environment–management interactions, and distribution shifts between development and deployment. Mechanism- and knowledge-guided learning, multimodal integration, and edge–cloud computing can improve generalization, interpretability, or deployability in specific settings, but their benefits remain context-dependent and cannot be inferred from within-dataset accuracy alone. These findings reveal a fundamental boundary between statistical prediction and reliable agricultural intelligence, particularly under limited labels, heterogeneous environments, uncertain observations, and resource-constrained deployment. Future research should therefore prioritize interoperable multimodal data infrastructures, mechanism-guided and uncertainty-aware learning, cross-environment validation, and scalable human–AI decision systems rather than isolated algorithmic advances. By integrating technological evolution with biological constraints and decision requirements, this review reframes crop growth AI as a transition from task-specific pattern recognition to context-aware decision intelligence and defines the conditions that govern its reliability, transferability, and agricultural value.
