State-dependent errors in a land surface model across biomes inferred from eddy covariance observations on multiple timescales

Characterization of state-dependent model biases in land surface models can highlight model deficiencies, and provide new insights into model development. In this study, artificial neural networks (ANNs) are used to estimate the state-dependent biases of a land surface model (ORCHIDEE: ORganising Carbon and Hydrology in Dynamic EcosystEms). To characterize state-dependent biases in ORCHIDEE, we use multi-year flux measurements made at 125 eddy covariance sites that cover 7 different plant functional types (PFTs) and 5 climate groups. We determine whether the state-dependent model biases in five flux variables (H: sensible heat, LE: latent heat, NEE: net ecosystem exchange, GPP: gross primary productivity and Reco: ecosystem respiration) are transferable within and between three different timescales (diurnal, seasonal–annual and interannual), and between sites (categorized by PFTs and climate groups). For each flux variable at each site, the spectral decomposition method (singular system analysis) was used to reconstruct time series on the three different timescales.At the site level, we found that the share of state-dependent model biases (hereafter called “error transferability”) is larger for seasonal–annual and interannual timescales than for the diurnal timescale, but little error transferability was found between timescales in all flux variables. Thus, performing model evaluations at multiple timescales is essential for diagnostics and future development. For all PFTs, climate groups and timescale components, the state-dependent model biases are found to be transferable between sites within the same PFT and climate group, suggesting that specific model developments and improvements based on specific eddy covariance sites can be used to enhance the model performance at other sites within the same PFT-climate group. This also supports the legitimacy of upscaling from the ecosystem scale of eddy covariance sites to the regional scale based on the similarity of PFT and climate group. However, the transferability of state-dependent model biases between PFTs or climate groups is not always found on the seasonal–annual and interannual timescales, which is contrary to transferability found on the diurnal timescale and the original time series.

► We investigate model bias transferability in both timescale and spatial domains. ► 125 eddy covariance sites on a global scale are used for this study. ► Model bias is more transferable on long timescales than the short one. ► Model bias is transferable between sites within the same PFT/climate group.