Revisiting PLUMBER: Why Do Simple Data-driven Models Outperform Modern Land Surface Models?

Monday, 14 December 2015: 15:10
3012 (Moscone West)
Bart Nijssen, University of Washington Seattle Campus, Seattle, WA, United States, Martyn P Clark, National Center for Atmospheric Research, Boulder, CO, United States, Ned Haughton, University of New South Wales, Climate Change Research Centre (CCRC), Sydney, Australia and Gab Abramowitz, University of New South Wales, Sydney, Australia
PLUMBER, a recent benchmarking study for the performance of land surface models (LSMs), demonstrated that simple data-driven models outperform modern LSMs at FLUXNET stations. Specifically, data-driven models out-performed LSMs in partitioning net radiation into turbulent heat fluxes over a wide range of performance criteria. The question is why. After all, LSMs combine process understanding with site information and might be expected to outperform simple data-driven models that are trained out-of-sample and that do not include an explicit representation of past states such as soil moisture or heat storage. In other words, the data-driven models have no explicit representation of memory, which we know to be important for land surface energy and moisture states.

Here, we revisit the PLUMBER results with the aim to understand why simple data-driven models outperform LSMs. First, we analyze the PLUMBER results to determine the conditions under which data-driven models outperform LSMs. We then use the Structure for Unifying Multiple Modeling Alternatives (SUMMA) to construct LSMs of varying complexity to relate model performance to process representation. SUMMA is a hydrologic modeling approach that enables a controlled and systematic analysis of alternative modeling options. Results are intended to identify development priorities for LSMs.