Estimation of regional net primary productivity (NPP) using a process-based ecosystem model: How important is the accuracy of climate data?

In this study we compared two different climate datasets, one from the National Center for Atmospheric Research (NCAR) of the USA and the other one from the Japan Weather Association (JWA), by inputting the climate data to the same ecosystem model, the Boreal Ecosystem Productivity Simulator (BEPS), to examine how the data quality of the climate input affected the estimate of NPP in 1998 on Hokkaido Island, Japan. The JWA dataset was derived from a high-density weather station network with adjacent stations about 17–21 km apart and served as our high quality climate data. The original NCAR dataset (about 0.9°×0.9°) was resampled at 1 km×1 km resolution using the bilinear interpolation method and served as our “poor” quality data source. We found that, on average, the NCAR dataset underestimated the 1998 NPP by 15% over the whole study area compared with the JWA dataset, by 21% in needle forest, by 18% in grass/cereal-cropland, by 14% in broadleaf forest, by 11% in broadleaf crops, and by 2% in shrubs. However, the NCAR dataset slightly overestimated NPP by 1% in savannas in comparison with the JWA dataset. o found that using the NCAR temperature data (daily maximum and minimum temperatures), specific humidity, and precipitation underestimated regional NPP by 17, 7, and 2%, respectively, compared with the JWA dataset. However, using the NCAR global solar radiation did not significantly change the estimate of NPP although the NCAR solar radiation was much greater than the JWA solar radiation. By using a new sensitivity analysis technique called the Extension of Fourier Amplitude Sensitivity Test (EFAST), we found that the required accuracy for the 1998 growing season was ±2.9 °C for daily maximum temperature, ±2.0 °C for daily minimum temperature, ±2.5 MJ/m2 per day for daily global solar radiation, and ±3.4 g/kg for daily specific humidity to estimate NPP with an overall confidence coefficient of 95% in the Hokkaido Island. The analytical approach developed in this study can also be applied to other input variables/parameters and other ecosystem models.