Improving the performance of remote sensing models for capturing intra- and inter-annual variations in daily GPP: an analysis using global FLUXNET tower data

Accurate and reliable estimates of gross primary productivity (GPP) are required for monitoring the global carbon cycle at different spatial and temporal scales. Because GPP displays high spatial and temporal variation, remote sensing plays a major role in producing gridded estimates of GPP across spatiotemporal scales. In this context, understanding the strengths and weaknesses of remote sensing-based models of GPP and improving their performance is a key contemporary scientific activity. We used measurements from 157 research sites (∼470 site-years) in the FLUXNET “La Thuile” data and compared the skills of 11 different remote sensing models in capturing intra- and inter-annual variations in daily GPP in seven different biomes. Results show that the models were able to capture significant intra-annual variation in GPP (Index of Agreement = 0.4–0.80) in all biomes. However, the models’ ability to track inter-annual variation in daily GPP was significantly weaker (IoA < 0.45). We examined whether the inclusion of different mechanisms that are missing in the models could improve their predictive power. The mechanisms included the effect of sub-daily variation in environmental variables on daily GPP, factoring-in differential rates of GPP conversion efficiency for direct and diffuse incident radiation, lagged effects of environmental variables, better representation of soil-moisture dynamics, and allowing spatial variation in model parameters. Our analyses suggest that the next generation remote sensing models need better representation of soil-moisture, but other mechanisms that have been found to influence GPP in site-level studies may not have significant bearing on model performance at continental and global scales. Understanding the relative controls of biotic vis-a-vis abiotic factors on GPP and accurately scaling up leaf level processes to the ecosystem scale are likely to be important for recognizing the limitations of remote sensing model and improving their formulation