Simulating Gross Primary Productivity By Assimilating Remote Sensing Data With A Two-leaf Light Use Efficiency Model

Terrestrial ecosystem carbon cycle is one of the hot spots in global change studies currently. It could not only have a significant impact on the global climate system, but also be closely linked with many material cycles. Terrestrial ecosystem productivity is one of the most important indicators to characterize terrestrial carbon cycle. Thus, it is of great importance to the research of carbon cycle. Gross Primary Productivity (GPP) is not only one of the major determinants of carbon exchange between the atmosphere and terrestrial ecosystems, but also an crucial gauge to describe plant activities and functions. Quantitative estimates of the spatial and temporal distribution of GPP at regional or global scales are of distinct significance for understanding the response of ecosystems to increases in atmospheric CO2and temperature and are thus central to policy-relevant decisions.This study takes Qianyanzhou subtropical evergreen coniferous forest flux site (QYZ) as the study site. With smoothed MODIS LAI and in situ measured meteorological data and flux data in2003-2005, a two leaf light use efficiency model (TL-LUE) on the basis of MOD17algorithm was developed. The TL-LUE model considered the differences in absorbing and using photosynthetically active radiation (PAR) of different leaves. The maximum light use efficiency (LUE) in the MOD17algorithm and TL-LUE model was calibrated using tower based data. The advantages of TL-LUE model for GPP simulations were analyzed at daily and8-day scales. Meanwhile, the smoothed MODIS LAI and LAI estimated by the TL-LUE model were assimilated by the Ensemble Kalman Filter (EnKF) to further improve GPP simulations. The main conclusions in this study could be drawn as follows: (1) With using in situ measured meteorological data, smoothed MODIS LAI and the calibrated maximum LUE (εmax) in2003-2005as inputs, GPP simulated by MOD17algorithm (GPPMOD) could effectively improve the serious underestimation of MODIS GPP product in the study site. GPPMOD had similar variations with measured GPP (GPPEC), with the determination coefficient (R2) above0.7and root mean square error (RMSE) below1.99g C m2d-1. The validation in2006also indicated that with calibrated εMAX, GPP estimated by the MOD17algorithm could have a good consistency with GPPEC.Therefore, MOD17algorithm was applicable to simulate GPP in subtropical coniferous forests in South China whenemax is properly determined and high quality of input data are used.(2) In order to improve GPP simulation, based on the MOD17algorithm, the TL-LUE model considered the difference in absorbing and using solar radiation by different leaves, and stratified the canopy into sunlit and shaded leaf groups to calculate GPP separately. In2003-2005, the R2of simulated GPP (GPPTL) and GPPEC was0.84-0.89, and RMSE was1.26-1.80g C m-2d-1. In comparison with the MOD17algorithm, the TL-LUE model reduced the sensitivity to sky clearness significantly, and could improve GPP simulation at daily scale than at8-day scale. The TL-LUE model remains the advantages of light use efficiency models which are simple in structure, less requirements for inputs. and easy parameterization and is able to simulate GPP more accurately. Therefore, the TL-LUE model could be applied to simulate GPP at regional or global scales by using remotely sensed data.(3) The calibrated maximum LUE of shaded leaves(εmsh) was distinctly larger than the calibrated εmax in the MOD17algorithm. It was about2times as much as εmax. However, the calibrated maximum LUE of sunlit leaves (εmsu) was considerable lower than the calibrated εmox. It was only about38%-53%of εmax.These findings confirm the hypothesis that the maximum LUE of different leaves within the canopy could differ significantly.(4) By assimilating MODIS LAI with LAI simulated by the TL-LUE model using the Ensemble Kalman Filter method, the consistency between simulated GPP and GPPEC was further improved, with the R2increasing by0-0.04and RMSE decreasing by0.02-0.39g C m-2d-1. The most significant improvement for GPP simulation was in winter and spring, owing to the effective correction of serious underestimation of MODIS LAI through data assimilation.