Evaluation and comparison of gross primary production estimates for the Northern Great Plains grasslands
Two spatially-explicit estimates of gross primary production (GPP) are available for the Northern Great Plains. An empirical piecewise regression (PWR) GPP model was developed from flux tower measurements to map carbon flux across the region. The Moderate Resolution Imaging Spectrometer (MODIS) GPP model is a process-based model that uses flux tower data to calibrate its parameters. Verification and comparison of the regional PWR GPP and the global MODIS GPP are important for the modeling of grassland carbon flux. This study compared GPP estimates from PWR and MODIS models with five towers in the grasslands. Among them, PWR GPP and MODIS GPP showed a good agreement with tower-based GPP at three towers. The global MODIS GPP, however, did not agree well with tower-based GPP at two other towers, probably because of the insensitivity of MODIS model to regional ecosystem and climate change and extreme soil moisture conditions. Cross-validation indicated that the PWR model is relatively robust for predicting regional grassland GPP. However, the PWR model should include a wide variety of flux tower data as the training data sets to obtain more accurate results.
In addition, GPP maps based on the PWR and MODIS models were compared for the entire region. In the northwest and south, PWR GPP was much higher than MODIS GPP. These areas were characterized by the higher water holding capacity with a lower proportion of C4 grasses in the northwest and a higher proportion of C4 grasses in the south. In the central and southeastern regions, PWR GPP was much lower than MODIS GPP under complicated conditions with generally mixed C3/C4 grasses. The analysis indicated that the global MODIS GPP model has some limitations on detecting moisture stress, which may have been caused by the facts that C3 and C4 grasses are not distinguished, water stress is driven by vapor pressure deficit (VPD) from coarse meteorological data, and MODIS land cover data are unable to differentiate the sub-pixel cropland components.