Using remote sensing to identify the peak of the growing season at globally-distributed flux sites: A comparison of models, sensors, and biomes

Abstract

The peak of the growing season (POS) is known to play an important role in regulating the interannual variability of terrestrial carbon sequestration. Recent research has developed several models for POS estimation; however, a comprehensive understanding of the predictive ability of these models remains lacking, especially taking various biomes and satellite data from different sensors into account. Using data from 54 eddy covariance (EC) flux sites (434 site-years in total) from the FLUXNET2015 dataset, we extracted POS from the normalized difference vegetation index (NDVI), derived from two sensors (MODIS and SPOT-VGT), using four different methods. We then compared the model outputs when data from the different sensors were used, and across different biomes. Our results show that the model predictions correlated weakly (R2 < 0.4) with the flux-based POS when multiple biomes were considered together. However, the performance of the models varied significantly between the models, the sensors that provided the data, and different biomes. Firstly, the more recently proposed methods did not perform as expected, and some of them performed even worse than the commonly used approach. Secondly, POS modeled from MODIS data performed slightly better than that from SPOT-VGT data. Thirdly, when the models are combined, they can reliably estimate POS for grasslands, deciduous broadleaf forests, and open shrublands, but not necessarily for other biomes. Lastly, our results indicate that NDVI-based POS is not a good proxy of flux-based POS. The study suggests that both biomes and sensor properties should be taken into account when estimating POS, and a rigorous validation is necessary before different models are implemented at regional, or larger scales. Therefore, this study provides insights that are helpful for improving our understanding of the impacts of algorithms, sensors, and biomes on model estimates of POS.