Multi-site land surface model optimization: An exploration of objective functions

• Published in: Agricultural and forest meteorology Vol. 182-183; pp. 168 - 176
• Main Authors: Fischer, Graciela R., Costa, Marcos H., Murta, Fabrício Z., Malhado, Ana C.M., Aguiar, Leonardo J.G., Ladle, Richard J.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15.12.2013
• Subjects: Adjustment measure; Amazonia; atmosphere; Biosphere model; Calibration; Ecosystems; Errors; Exchange; Land; LBA-DMIP; Mathematical models; net ecosystem exchange; Pasture; pastures; Time; Amazonia; South America, Amazonia; Biosphere model; Calibration; LBA-DMIP; Pasture; Adjustment measure
• ISSN: 0168-1923; 1873-2240
• DOI: 10.1016/j.agrformet.2012.11.021

► Twenty objective functions are evaluated for multi-site model optimization. ► Multi-site calibration is possible and produces consistent results across sites. ► The choice of objective function should be based on the intended use of the model. Land surface/ecosystem models are important tools for understanding the dynamic interactions between land surface and the atmosphere. However, to be effective these models must be carefully calibrated to accurately represent ecosystem processes. Generally, such models are calibrated for one site and then run with the same set of calibrated parameters, either for other sites or for a whole region with the same plant functional type. Here we investigate an alternative approach to the challenge of calibration. We perform multi-site calibration of net ecosystem exchange for two pasture sites in Amazonia. Twenty different objective functions (five adjustment measures subject to four calibration options) are evaluated to investigate the consistency and sensitivity of the results in a multi-site model calibration. Our results indicate that, with some restrictions regarding the choice of objective function, multi-site calibration is possible and produces consistent results across sites. Ultimately, the choice of objective function should be based on the intended use of the model. We recommend that the site-weighted method using mean absolute error as objective function should be used for shorter time scales and the site-weighted maximum bias error as objective function is better for longer time scales.