Discharge and soil loss

The effect of input data source showed to be very important when using the event based LISEM model to simulate catchment discharge (surface runoff) and soil loss (paper IV). Input data derived from the national soil survey database (“Generic run”) resulted in a peak discharge that was almost 400 l/s higher than the peak discharge simulated using locally measured data (“Mean run” and “Stochastic runs”), time to peak was 130 minutes earlier and the soil loss was five times higher (Figures 4 and 5). In most circumstances measured data will not be available, and the only option is to use the soil survey database. It is therefore important to take uncertainty into account.

The value of Ks, derived from the basic input data using a pedotransfer function, was especially important in explaining the differences, as LISEM is highly sensitive to this parameter (De Roo et al., 1996b; Stolte et al., 2003). Less surface runoff simulated by the Mean run and Stochastic runs corresponded with higher area-weighted mean Ks (86 and 92 cm d-1, respectively) on arable land in these runs than in the Generic run (59 cm d-1). The predicted Ks will strongly depend on the basic soil properties as long as PTFs are used, emphasizing the importance of having access to adequate basic data and PTFs. Ks poses a large problem in modeling because it is often highly variable in time and space, it has a high degree of random variability (often resulting from macroporosity) and it is very difficult to measure correctly. Some authors have suggested that Ks is best represented as a calibration parameter (e.g. Davis et al., 1999), but this is inappropriate in distributed modeling, where the spatial distribution of Ks is important. In our study a “compromise” solution was chosen, where the PTF predicted Ks of all clay soils (because of macroporosity) was multiplied with a single calibration factor to get a good match between the measured and simulated hydrograph.

More measured Ks data are required to come up with a more sophisticated and realistic approach.

The main reason for more soil loss simulated by the Generic run seemed to be that considerably more surface runoff was simulated by this approach. Differences in aggregate stability and cohesion were too small to explain the difference.

The two approaches of assigning locally measured input data, i.e. using a mean value on one hand and a stochastic distribution on the other hand, did not result in large differences in simulated discharge and soil loss. The variability in model output for the realizations in the

stochastic approach was negligible, especially for simulated surface runoff. The possibly small gain in precision by using the stochastic approach instead of the mean value approach cannot justify the extra effort made in input data generation, model runs and processing of results for multiple realizations. This needs to be verified also for other situations (surface cover, event size, season). Both of these approaches to assign measured data used the soil map as basis for spatial distribution of soil properties. This study should be followed up with alternative approaches where the spatial correlation of the locally measured data is retained.

This might influence the result due to interactions between soil and terrain.

Running the model for both a low risk situation (crop covered surface) and a high risk situation (bare soil with freeze-thaw induced changes in aggregate stability and cohesion) showed that the variability in model output for the stochastic approach was similar in both cases, i.e., the uncertainty did not appear to depend on the erosion risk. Comparison of absolute and relative differences between the two risk situations and the three input data approaches showed that the uncertainty related to input data could result in larger differences between runs with different input data source than between runs with the same input data source but extreme differences in erosion risk. Effects of removing the crop cover and decreasing the structural stability varied between the input data approaches.

Figure 4. Simulated surface runoff for the Generic run, Mean run and 50 Stochastic runs, together with rainfall intensity.

Figure 5. Simulated soil loss for the different input data approaches, summer conditions and “worst case” winter conditions. The bars shown for the Stochastic and Combined approach represent the mean of the 50 realizations, while the whiskers represent minimum and maximum values.

This study clearly showed that inadequate choice of input data sources can

significantly underestimate or overestimate surface discharge, soil loss and consequently the effect of measures to reduce soil erosion. Clearly, this is only a first step towards quantifying uncertainty in simulated responses of catchments in Norway. Other studies dealing with input data uncertainty in modeling show that the effects of input data source, variability and

uncertainty differ between models, type of response, spatial and temporal resolution, thresholds for certain parameters, event size, etc. (e.g. Merz and Plate, 1997; Vachaud and Chen, 2002; Lindahl et al., 2005).

Unfortunately, we cannot find out which of the approaches used in this study produced the most realistic results at the main catchment scale, because only total discharge is measured at the catchment outlet. The total discharge is a mix of surface runoff, subsurface drainage water (both macropore and matrix flow), and groundwater flow/ baseflow. Investigations of uncertainty in relation to measured response should be investigated on a sub-catchment scale, with available precipitation and surface runoff data with high temporal resolution (sub-hourly or even sub-minute). The sub-catchment within the Skuterud catchment, successfully used for

calibration of LISEM, is the only site in Norway that can be used for this specific purpose. It is strongly recommended that more such experimental sites are established for more detailed research on effects of variability and uncertainty in input data on hydrological processes, soil and nutrient loss.

Lack of validation data can be an equally important problem in hydrological modeling as lack of input data – we need to have some confidence that the models produce realistic responses in order to use them in risk and scenario analyses. Spatially distributed models like LISEM have the advantage that they can be used to predict not only the magnitude of loss from a catchment, but also where the problems are largest and where it will be most beneficial to invest in measures. At the same time, it is more difficult and costly to obtain appropriate input data and validation data. A study by Takken et al. (1999) clearly showed that it is not possible to evaluate the performance of spatially distributed erosion models by using catchment outlet data alone; such models’ behavior can only be understood if they are evaluated using spatially distributed data. Poor reproduction of spatial erosion patterns was also found by Jetten et al. (2003) and Hessel et al. (2003). Hessel et al. (2006) has pointed out several reasons for overpredicted soil loss and poor match between simulated and observed erosion patterns: 1) the difficulty of obtaining enough accurate data to run the model, 2) the difficulty of obtaining accurate data for validation, both at the outlet and spatially, 3) the model could not deal correctly with complex events, i.e. those having double rainfall peaks, and 4) the model could not deal with events in which throughflow or baseflow played a role (those processes are not simulated). De Roo (1998) relates the disappointing results of spatial models in the uncertainty involved in estimating and measuring the large number of input variables at a catchment scale. Presently model performance in predicting spatial erosion patterns cannot be carried out for the Skuterud catchment or any other place in Norway, due to lack of both erosion pattern observations and spatially distributed model input data. It is recommended that future research on soil erosion puts strong emphasis on prediction of spatial patterns, because, as stated by Jetten et al. (2003): “It is much more cost effective to over-dimension an erosion control measure than to put it in the wrong spot”. Sometimes the benefit from using a complex, process based model is counteracted by the increasing error due to uncertainty on the soil information for such models (Van Rompaey and Govers, 2002;

Schoups and Hopmans, 2006). Thus, simpler (e.g. lumped, regression based) models may actually outperform the more complex models.