Quantifying uncertainty in remote sensing of plant nitrogen status: A three–level meta–analysis and Bayesian network approach

Accurate and timely monitoring of plant nitrogen (N) status is essential for understanding vegetation productivity, soil health, and climate feedbacks, yet large–scale assessment remains challenging due to ecosystem heterogeneity and human disturbances. Remote sensing provides an effective approach for monitoring vegetation N across different spatial and temporal scales, but estimation accuracy varies widely among studies. The objective of this review was to identify key factors driving variability in remote sensing–based estimates of plant N status and their interactions. A three–level meta–analysis with Bayesian network modeling was combined to synthesize 162 studies across 261 locations worldwide, covering 11 preprocessing methods, 28 feature selection approaches, and 27 estimation models. The results showed that preprocessing, feature selection, and estimation models were the main determinants of model performance. Spectral preprocessing improved accuracy compared with the original reflectance, underscoring the importance of noise reduction and spectral correction. Embedded feature selection performed the best (R 2 = 0.72) among the tested methods. Among estimation models, hybrid models (R 2 = 0.75) and nonlinear machine–learning models (R 2 = 0.75) achieved the highest accuracy. Sensor type and observation conditions also influenced performance. Hyperspectral and hybrid sensors generally outperformed RGB and multispectral systems, while ground–based or multi–platform observations yielded higher accuracy. The influence of the Bayesian network further revealed that higher estimation accuracy was associated with context-specific combinations of methodological factors rather than single techniques. For example, for leaf–scale N estimation on a mass basis, combining wrapper—based feature selection, such as the Genetic Algorithm, with linear nonparametric regression models, such as partial least squares regression, yielded a 39% probability of achieving R 2 ≥ 0.8. For canopy–scale N estimation on an area basis, combining wrapper–based methods with nonlinear nonparametric regression models, such as Gaussian process regression, yielded a similar probability (39%) of achieving R 2 ≥ 0.8, highlighting the importance of matching modeling strategies to observational scales and N metrics. These patterns suggested that, in vegetation N status estimation, predictive performance depended on specific contextual settings, highlighting the importance of appropriately designing methodological combinations across conditions rather than focusing on a single methodological factor. Overall, the integrated framework combining three-level meta-analysis with Bayesian network modeling enhanced conventional meta-analysis by revealing not only general methodological patterns, but also context-dependent interactions and methodological combinations for higher estimation accuracy for vegetation N status.