Print PageWe evaluated the predictive ability of a one-dimensional coupled hydrodynamic biogeochemical model across multiple temporal scales using wavelet analysis and traditional goodness-of-fit metrics. High-frequency in situ automated sensor data and long-term manual observational data from Lake Mendota, Wisconsin, USA, were used to parameterize, calibrate, and evaluate model predictions. We focused specifically on short-term (< 1 month) predictions of phytoplankton biomass over one season.
Results/Conclusions
Traditional goodness-of-fit metrics indicated more accurate prediction of physics than chemical or biological variables in the time domain. This was confirmed by wavelet analysis in both the time and frequency domains. For temperature, predicted and observed global wavelet spectra were closely related, while observed dissolved oxygen and chlorophyll-a fluorescence spectral characteristics were not reproduced by the model for key time scales, indicating that processes not modelled may be important drivers of the observed signal. To determine the influence of exogenous drivers and starting conditions on system dynamics, we simulated the response of dominant phytoplankton groups to different nutrient and water temperature scenarios. We found that the initial conditions of water column phosphorus concentration was more important to the timing and magnitude phytoplankton response than nitrogen concentration or initial water column temperature. Dynamics of simulated variables did not change with starting conditions or changes in temperature or phosphorus loading. Although the magnitude and timing of physical and biological changes can be simulated adequately through calibration, time-scale specific dynamics, for example short-term cycles, are difficult to reproduce. Wavelet transforms and diverse observational data provide for model evaluation techniques that are complementary to traditional goodness-of-fit metrics, and are particularly well suited for assessment of temporal and spatial heterogeneity when coupled to high-frequency data from automated in situ or remote sensing platforms.
