Distributed modeling of storm flow generation in an Amazonian rain forest catchment: Effects of model parameterization

We describe a process‐based storm flow generation model, Topog_SBM consisting of a simple bucket model for soil water accounting, a one‐dimensional kinematic wave overland flow scheme, and a contour‐based element network for routing surface and subsurface flows. Aside from topographic data and rainfall the model has only six input parameters: soil depth (z), saturated hydraulic conductivity at the soil surface (K0), the rate of decay in K0 with depth (m), the Manning surface roughness parameter (n), the maximum (saturated) soil water content (θs), and the minimum (residual) soil water content (θr). However, the model is fully distributed, so these values can vary in magnitude across space. The model was applied to La Cuenca, a very small rainforest catchment in western Amazonia that has been well characterized in several hydrometric and hydrochemical investigations. Total runoff, peak runoff, time of rise, and lag time were predicted for 34 events of varying magnitudes and antecedent moisture conditions. We compared results for eight different model parameterizations or “sets”; four of these were freely calibrated to yield the best possible model fit to runoff data, whereas the other four were constrained (in various ways) by the use of actual K0 data gathered for the catchment. The eight sets were calibrated on either one of three events or on the three events jointly to illustrate the importance of calibration event selection on model performance. Model performance was evaluated by comparing observed and predicted (1) storm flow hydrograph attributes and (2) spatiotemporal patterns of overland flow occurrence across the catchment. The model generally predicted the right amount of runoff but usually underpredicted the peak runoff rate and overpredicted the time of rise. The “best” parameterization could credibly predict hydrographs for only about half of the events. Significant, and sometimes gross, errors were encountered for about one fourth of the events modeled, raising concerns in our minds about the a priori simulation of events that diverge too far from the conditions that the model was calibrated for. For the best parameterization we were able to predict an overland flow frequency distribution that accorded with field observations, though the model almost always overpredicted the spatial extent of overland flow. We concluded that model performance for the La Cuenca conditions could be enhanced by adding a “fast” subsurface flow pathway and/or by m