Observing and modeling dynamics in terrestrial gross primary productivity and phenology from remote sensing: An assessment using in-situ measurements

Terrestrial gross primary productivity (GPP) is the largest and most variable component of the carbon cycle and is strongly influenced by phenology. Realistic characterization of spatio-temporal variation in GPP and phenology is therefore crucial for understanding dynamics in the global carbon cycle. In the last two decades, remote sensing has become a widely-used tool for this purpose. However, no study has comprehensively examined how well remote sensing models capture spatiotemporal patterns in GPP, and validation of remote sensing-based phenology models is limited. Using in-situ data from 144 eddy covariance towers located in all major biomes, I assessed the ability of 10 remote sensing-based methods to capture spatio-temporal variation in GPP at annual and seasonal scales. The models are based on different hypotheses regarding ecophysiological controls on GPP and span a range of structural and computational complexity. The results lead to four main conclusions: (i) at annual time scale, models were more successful capturing spatial variability than temporal variability; (ii) at seasonal scale, models were more successful in capturing average seasonal variability than interannual variability; (iii) simpler models performed as well or better than complex models; and (iv) models that were best at explaining seasonal variability in GPP were different from those that were best able to explain variability in annual scale GPP.;Seasonal phenology of vegetation follows bounded growth and decay, and is widely modeled using growth functions. However, the specific form of the growth function affects how phenological dynamics are represented in ecosystem and remote sensing-base models. To examine this, four different growth functions (the logistic, Gompertz, Mirror-Gompertz and Richards function) were assessed using remotely sensed and in-situ data collected at several deciduous forest sites. All of the growth functions provided good statistical representation of in-situ and remote sensing time series. However, the Richards function captured observed asymmetric dynamics that were not captured by the other functions. The timing of key phenophase transitions derived using the Richards function therefore agreed best with observations. This suggests that ecosystem models and remote-sensing algorithms would benefit from using the Richards function to represent phenological dynamics.