Assessing Ecosystem Model Performance in Semiarid Systems

In ecosystem process modelling, comparing outputs to benchmark datasets observed in the field is an important way to validate models, allowing the modelling community to track model performance over time and compare models at specific sites. Multi-model comparison projects as well as models themselves have largely been focused on temperate forests and similar biomes. Semiarid regions, on the other hand, are underrepresented in land surface and ecosystem modelling efforts, and yet will be disproportionately impacted by disturbances such as climate change due to their sensitivity to changes in the water balance. Benchmarking models at semiarid sites is an important step in assessing and improving models' suitability for predicting the impact of disturbance on semiarid ecosystems. In this study, several ecosystem models were compared at a semiarid grassland in southwestern Arizona using PEcAn, or the Predictive Ecosystem Analyzer, an open-source eco-informatics toolbox ideal for creating the repeatable model workflows necessary for benchmarking. Models included SIPNET, DALEC, JULES, ED2, GDAY, LPJ-GUESS, MAESPA, CLM, CABLE, and FATES. Comparison between model output and benchmarks such as net ecosystem exchange (NEE) tended to produce high root mean square error and low correlation coefficients, reflecting poor simulation of seasonality and the tendency for models to create much higher carbon sources than observed. These results indicate that ecosystem models do not currently adequately represent semiarid ecosystem processes.