Land Surface Model Underperformance Tied to Specific Meteorological Conditions

Abstract. The exchange of carbon, water, and energy fluxes between the land and the atmosphere plays a vital role in shaping our understanding of global change and how climate change will affect extreme events. Yet our understanding of the theory of this surface-atmosphere exchange, represented via land surface models, continues to be limited, highlighted by marked biases in model-data benchmarking exercises. Here, we leveraged the PLUMBER2 dataset of observations and model simulations of terrestrial fluxes from 153 international eddy-covariance sites to identify the meteorological conditions under which land surface models are performing worse than a priori expectations. By defining performance relative to three sophisticated out-of-sample empirical models, we generated a lower bound of performance in turbulent flux prediction that can be achieved with the input information available to the land surface models during testing at flux tower sites. We found that land surface model (LSM) performance relative to empirical models is worse at boundary conditions – that is, LSMs underperform in timesteps where the meteorological conditions consist of coinciding relative extreme values. Conversely, LSMs perform much better under "typical" conditions within the centre of the meteorological variable distributions. Constraining analysis to exclude the boundary conditions results in the LSMs outperforming strong empirical benchmarks. Encouragingly, we show that refinement of the performance of land surface models in these boundary conditions, consisting of only 12 % to 31 % of time steps, would see large improvements (22 % to 114 %) in an aggregated performance metric. Precise targeting of model development towards these meteorological boundary conditions offers a fruitful avenue to focus model development, ensuring future improvements have the greatest impact.