Reducing impacts of systematic errors in the observation data on inversing ecosystem model parameters using different normalization methods

Abstract. Modeling ecosystem carbon cycle on the regional and global scales is crucial to the prediction of future global atmospheric CO₂ concentration and thus global temperature which features large uncertainties due mainly to the limitations in our knowledge and in the climate and ecosystem models. There is a growing body of research on parameter estimation against available carbon measurements to reduce model prediction uncertainty at regional and global scales. However, the systematic errors with the observation data have rarely been investigated in the optimization procedures in previous studies. In this study, we examined the feasibility of reducing the impact of systematic errors on parameter estimation using normalization methods, and evaluated the effectiveness of three normalization methods (i.e. maximum normalization, min-max normalization, and z-score normalization) on inversing key parameters, for example the maximum carboxylation rate (V_cmax,25) at a reference temperature of 25°C, in a process-based ecosystem model for deciduous needle-leaf forests in northern China constrained by the leaf area index (LAI) data. The LAI data used for parameter estimation were composed of the model output LAI (truth) and various designated systematic errors and random errors. We found that the estimation of V_cmax,25 could be severely biased with the composite LAI if no normalization was taken. Compared with the maximum normalization and the min-max normalization methods, the z-score normalization method was the most robust in reducing the impact of systematic errors on parameter estimation. The most probable values of estimated V_cmax,25 inversed by the z-score normalized LAI data were consistent with the true parameter values as in the model inputs though the estimation uncertainty increased with the magnitudes of random errors in the observations. We concluded that the z-score normalization method should be applied to the observed or measured data to improve model parameter estimation, especially when the potential errors in the constraining (observation) datasets are unknown.