资源环境科技发展态势分析平台

The related challenges of predictions in ungauged basins and predictions in ungauged climates point to the need to develop environmental models that are transferable across both space and time. Hydrologic modeling has historically focused on modelling one or only a few basins using highly parameterized conceptual or physically based models. However, model parameters and structures have been shown to change significantly when calibrated to new basins or time periods, suggesting that model complexity and model transferability may be antithetical. Empirical space-for-time models provide a framework within which to assess model transferability and any tradeoff with model complexity. Using 497 SNOTEL sites in the western U.S., we develop space-for-time models of April 1 SWE and Snow Residence Time based on mean winter temperature and cumulative winter precipitation. The transferability of the models to new conditions (in both space and time) is assessed using non-random cross-validation tests with consideration of the influence of model complexity on transferability. As others have noted, the algorithmic empirical models transfer best when minimal extrapolation in input variables is required. Temporal split-sample validations use pseudoreplicated samples, resulting in the selection of overly complex models, which has implications for the design of hydrologic model validation tests. Finally, we show that low to moderate complexity models transfer most successfully to new conditions in space and time, providing empirical confirmation of the parsimony principal.

Plain Language Summary A challenge for environmental modeling and prediction is to create models that work everywhere. This includes places where we don't have data and it also includes the future. An additional challenge is choosing an appropriate model for these new conditions.To select a model, we built snow models of varying complexity and evaluated how well they predicted snowpack in new locations and time periods. The models performed well when applied to new time periods since certain environmental variables remained constant over time, including shading and solar radiation, resulting in the time samples being statistically dependent. For these applications, validation tended to select overly complex models. When applied to new locations, the models provided good predictions as long as the conditions in the new location were not drastically different from the conditions the model was trained on. Finally, we found that simple to moderate-complexity models did better than complex models at predicting in new conditions.