Disclaimer: This material is being kept online for historical purposes. Though accurate at the time of publication, it is no longer being updated. The page may contain broken links or outdated information, and parts may not function in current web browsers. Visit https://espo.nasa.gov for information about our current projects.
A coupled Bayesian model selection and geostatistical regression modeling approach is adopted for empirical analysis of gross primary productivity (GPP) at six AmeriFlux sites, including the Kennedy Space Center Scrub Oak, Vaira Ranch, Tonzi Ranch, Blodgett Forest, Morgan Monroe State Forest, and Harvard Forest sites. The analysis is performed at a continuum of temporal scales ranging from daily to monthly, for a period of seven years. A total of 10 covariates representing environmental stimuli and indices of plant physiology are considered in explaining variations in GPP. Similarly to other statistical methods, the presented approach estimates regression coefficients and uncertainties associated with the covariates in a selected regression model. Unlike traditional regression methods, however, the approach also estimates the uncertainty associated with the selection of a single “best” model of GPP. In addition, the approach provides an enhanced understanding of how the importance of specific covariates changes with the examined timescale (i.e. temporal resolution). An examination of changes in the importance of specific covariates across timescales reveals thresholds above or below which covariates become important in explaining GPP. Results indicate that most sites (especially those with a stronger seasonal cycle) exhibit at least one prominent scaling threshold between the daily and 20-day temporal scales. This demonstrates that environmental variables that explain GPP at synoptic scales are different from those that capture its seasonality. At shorter time scales, radiation, temperature, and vapor pressure deficit exert the most significant influence on GPP at most examined sites. At coarser time scales, however, the importance of these covariates in explaining GPP declines. Overall, unique best models are identified at most sites at the daily scale, whereas multiple competing models are identified at longer time scales.