Earth Observation-Driven Inference for Level-4 Terrestrial Products: Process-Based, Machine Learning, and Hybrid Frameworks

Satellite Earth observation (EO) provides global, repeated measurements of vegetation structure and function, yet many variables of primary ecological and societal relevance, such as gross primary productivity (GPP), evapotranspiration (ET), biomass, and crop yield, are not directly observable. These Level-4 (L4) variables are instead inferred through model-based estimation systems that integrate EO data with models describing system dynamics, observation operators, and uncertainty, and should therefore be interpreted as model-conditioned estimates rather than direct observations. This review synthesizes three complementary approaches for L4 inference: (1) process-based data assimilation, (2) machine learning (ML), and (3) hybrid process–ML methods, corresponding to explicit, implicit, and integrated inference frameworks. We compare these approaches in terms of dynamical consistency, interpretability, scalability, generalization, and uncertainty representation, and examine how EO-specific constraints, such as scale mismatch, nonlinear observation operators, and temporal sampling, limit identifiability and shape inference outcomes. We show that the reliability of L4 products is governed not by EO data quality alone, but by the overall coherence of the full observation–retrieval–model–inference chain. Process-based frameworks provide physically consistent, uncertainty-aware state–space formulations but remain computationally demanding and sensitive to structural error. ML methods offer scalability and flexibility but are constrained by training-domain dependence, weak dynamical consistency, and limited physical interpretability of predictive uncertainty. Hybrid frameworks provide a pathway to reconcile these trade-offs, provided that compatibility between learned and mechanistic components is preserved. Advancing EO-based monitoring therefore requires a transition from standalone products toward coherent, scale-aware, and uncertainty-aware inference systems that explicitly integrate observation physics, model structure, and uncertainty across the full processing chain, and that support reproducible, operational deployment in cloud-native environments.