Climate-driven uncertainties in modeling terrestrial ecosystem net primary productivity in China

- Uncertainty of process-based model in regional scale has rarely been assessed.
- We quantified the uncertainty of simulated regional NPP induced by climate input dataset.
- Uncertainty of simulated NPP ranged from 3.53 to 3.96 Pg C induced by four climate input dataset.
- Station density higher than 1.0 s/104 km2 could meet the accuracy requirement of 0.1° spatial resolution simulation.
- The two reanalysis dataset had larger uncertainty in most regions.

Evaluating the uncertainties in regional/global carbon flux estimates is essential for better understanding of terrestrial carbon dynamics. At the regional scale, climate input data is an important source of model simulation uncertainty. In this study, a process-based ecosystem model, CEVSA, was run driven by four climate input datasets during 1980-2004, i.e., climate input datasets interpolated from 756 (756s) and 2400 weather stations (2400s), the NCEP/NCAR and Princeton reanalysis datasets. We used the 2400s dataset as the reference because it was derived from high density weather station interpolation. The simulated Net Primary Productivity (NPP) based on interpolated climate data from the 756s and the two reanalysis datasets were compared with that from the 2400s dataset. Then, we quantified the uncertainty of model simulations at regional-scale caused by climate input data, and evaluated the performance of different climate datasets across different eco-regions. Our results suggest that the 756s, Princeton and NCEP/NCAR reanalysis datasets overestimated the 25-year mean annual temperature by 7.66%-12.25% and the precipitation by 2.83%-8.43%, respectively; accordingly, the simulated NPP ranged from 3.53 to 3.96 Pg C, 6% to 12% higher than the reference over the entire China. The 756s and the two reanalysis datasets captured well the trend and interannual variations of annual NPP during the study period, but showed systematic errors in the total amount of NPP compared with the 2400s dataset. To increase the station density in the eco-regions with a station density greater than 1.0 station per 104 km2 (1.0 s/104 km2) would not decrease the uncertainty for model simulation at a 0.1° spatial resolution. The NCEP/NCAR and Princeton reanalysis datasets showed larger uncertainties in most eco-regions compared with the interpolated datasets. Our results also suggest that the accuracy of the NCEP/NCAR reanalysis data should be further improved in most eco-regions. On Qinghai-Tibet Plateau and in northwestern China, all four climate input datasets had relatively lower accuracy due to the limited observation data available. Future work should further evaluate the simulated NPP against observations and quantify the accuracy of driving climate data to decrease the uncertainty of model simulations at the regional scale.