1. Introduction
The two basic snowmelt approaches generally used in hydrologic modeling are energy balance models and temperature-index models [
1]. Temperature-index models are widely used in hydrologic studies because of the models’ performance and simplicity as well as the wide availability of temperature data [
2,
3,
4].
In temperature-index based runoff models such as SWAT, melt rates only vary as a function of elevation, because of the air temperature gradient [
5]. To overcome this drawback, a modified snow process was incorporated into SWAT in order to consider spatial variation of snowmelt/accumulation parameters by elevation band across subbasins. Previous studies using SWAT kept these parameters constant for an entire basin [
6,
7,
8,
9,
10]. While this method was successful in simulation of snowmelt flow, simulation of runoff from glaciered watersheds demands a distributed model for distinguishing seasonal snow from glacier. The new approach allows for the separation of seasonal snowmelt from glaciers’ melt based on the spatial variability of associated melt parameters.
The SWAT model has been applied worldwide, and its hydrologic components have been successfully tested in cases where streamflow was predominantly generated from rainfall events [
11,
12,
13,
14,
15]. The model less frequently has been applied in mountainous watersheds, and a few recent studies have been conducted to test and improve SWAT’s snow hydrology component. Fontaine et al. [
16] incorporated the elevation bands method with SWAT’s original snowmelt algorithm (temperature-index model), which improved the Nash–Sutcliffe coefficient of monthly runoff simulation from −0.70 to 0.86 in the 4999-km
2 Rocky Mountain Basin in Wyoming. Debele and Srinivasan [
4] incorporated a modified version of SNOW17 [
17] into SWAT and compared its performance with the temperature-index model in three watersheds (ranging from 22.28 to 7106.82 km
2), the results of which showed that the temperature-index model performed better than the SNOW17 model. Debele et al. [
18] incorporated the distributed process-based energy budget SNOWEB in the pixel and elevation band scales into SWAT and compared its performance with the temperature-index model. In this method, it was assumed that solar radiation varies not only with latitude and altitude of subbasins, which is applied in the current version of SWAT, but also with land surface inclinations (aspect and slope). The temperature-index based snowmelt computation method had overall Nash-Sutcliffe model efficiency coefficient ranging from 0.49 to 0.73 in simulation of monthly streamflow while the energy budget based approach had Nash-Sutcliffe efficiency coefficient ranging from 0.33 to 0.59. Zhang et al. [
8] applied SNOW17 in SWAT at the pixel scale. The SWAT model with temperature-index plus elevation bands performed as well as the SWAT model with SNOW17. Rahman et al. [
19] applied the modified temperature-index method [
5] to simulate the runoff from fully and partially glaciated subbasins of the Rhone River Basin. The applied method showed improvement in simulated streamflow from the main outlet of the basin even though the snowmelt parameters were lumped across the basin.
Distributed, process-based energy budget models have been tested in SWAT, but we focused on another component of frequently used glacier/snowmelt models. A major difference between SWAT melt processes and the melt routine of hydroglacial models arises from the associated melt parameter distribution (melt factor). In previous studies using SWAT, associated snowmelt parameters were considered to be uniform within each subbasin, but in the hydroglacial models, melt factors are spatially variable in pixel or band scales. A common approach is to assign two different melt factors to ice and snow. This enables the user to treat seasonal snow and glaciers differently; melt factors are generally reported to be higher for glaciers and lower for snow [
20]. The applied approach in this study separates seasonal snow from glaciers by setting the accumulation/melt parameters, including maximum and minimum melt factors, melt lag factor, melt temperature and snow fall temperature for the elevation bands of each subbasin. The details of the proposed approach are available at Omani et al. [
21] for three benchmark glaciers in Asia and Europe. The objective of this study was to extend the applied method to macro-scale river basins that are global in coverage and vary in climatic condition. Two snow process algorithms (lumped and distributed) were examined for their ability to simulate flow in five river basins that provide global assessment and feature contrasts in climatic conditions.
3. Results and Discussion
Table 3,
Table 4,
Table 5,
Table 6 and
Table 7 show an average of adjusted values and range of the calibration parameters throughout the river basins for lumped and modified snow algorithms. The symbol “s” in the tables stands for glacier-free elevation bands or snowy elevation bands and “g” stands for glaciered elevation bands over the ELA
0 altitudes. The calibration parameter values are averaged for the elevation bands and subbasins and the parameter range shows minimum and maximum parameter values based on elevation band-subbasin scale throughout the entire river basin. Mean SMFMX (Maximum melt factor) throughout all river basins is between 2.80 and 5.38, and 3.30 to 7.71, for snow and glacier, respectively. TIMP (Snow temperature lag factor) values are lower in the high elevation bands where the glaciers exist and range between 0.010 to 1.000 with average values between 0.025 for high elevation bands and 0.740 for low elevation bands. It can be seen from
Table 3 that the average values for SMFMX and SMFMN (Minimum melt factor) for the Narayani River Basin are larger than the obtained values for Rhone River Basin (
Table 4) as also was indicated by Kayastha et al. [
46].
Figure 4 shows how the simulated monthly flow and mean monthly flow fit the observed mean monthly flow cycle using the modified algorithm for some of the selected reaches. This degree of accuracy in simulation of the seasonal pattern of monthly flow is only achievable by adjusting melt parameters based on elevation bands. As demonstrated in
Figure 5, SMFMX changes in lower elevation bands had more influence on the rising limb of the flow curve, whereas the descending limb was more sensitive to SMFMX changes in the higher elevation bands (glaciered areas). This indicates that the late spring flow is under the influence of snowmelt at lower elevations and that late summer flow is controlled by glacier melt at higher elevations. This can also be investigated by analyzing the melting lag time at the elevation bands. Seasonal melting from October 1998 to October 1999 from each of the elevation bands of a typical subbasin with a wide range of elevations is presented in
Figure 6a. Melting for glacier-free elevation bands with seasonal snow cover are presented in
Figure 6b for high elevation bands with permanent snow cover or glacier. In
Figure 6a, the snowpack in elevation bands 1, 2, 3, 4, and 5 has completely vanished by mid-August, whereas in
Figure 6b the glacier ablation reaches its peak in August and continues through October. The melt was decreased at lower altitudes by decreasing the SMTMP (snow melt base temperature) and SFTMP (snowfall temperature) at lower elevation bands.
Calibration and validation results for all river basins are given in
Table 8 and
Table 9. In the watersheds with highly glaciered drainage areas (Reaches 2 and 23) in the Rhone River Basin, PBIAS was improved when the modified snow algorithm was used instead of lumping all of the parameters for entire watershed area Both RSR and NSE also were improved using the modified model, although the improvement is more significant in simulation of magnitude of flow. There is negligible improvement in the simulated flow from watersheds with small areas of glacier (Reach 4) when using the modified snow algorithm in comparison to the lumped snow algorithm.
For Narayani River Basin, PBIAS and NSE values in
Table 8 reveal similar model performance by both lumped and modified snow algorithm methods, possibly as a result of the diminishing influence of glacier/snowmelt on runoff resulting from coincident summer monsoon precipitation in the Narayani River Basin. This monsoonal influence leads to accurate prediction of seasonal flow by rainfall–runoff models without explicitly considering the snow hydrology process. In
Table 8, positive PBIAS in simulation of flow magnitude from all gauge stations (except Reach 96) means that the model has underpredicted the volume of flow. This systematic bias might be due to an underestimation of the NCEP reanalysis spring/summer precipitation (May–September) in the Narayani River basin of 20%.
The modified snow algorithm performed very well in simulation of monthly flow from the glaciered areas of the Vakhsh River Basin with RSR smaller than 0.5, PBIAS smaller than 10 percent and NSE greater than 0.70. Major difficulties in simulation of monthly flow from the Vakhsh River Basin were a lack of data for calibration and the low quality of the climate data, especially in the northern half of the river basin. The short calibration period of 2.5 years for Reach 72 was not long enough to capture the long-term variability of the monthly flow. The simulated mean monthly flow volume and cycle from the glaciated areas (drainage area to Reaches 72 and 133) by the modified snow algorithm during the summer months is in better agreement with the observed data in comparison to the mean monthly flow simulated using the lumped model for Reach 133 (
Figure 4). At the main outlet (Reach 109), only the monthly flow variation shows improvement. Consequently, for an entire river basin it cannot be concluded that the modified model has better performance than the lumped model. Comparison between observed and simulated monthly flow from the main outlet of the basin during the validation period indicated that there was good agreement, verified by RSR, PBIAS, and NSE of 0.52, −8.15, and 0.73, respectively.
Among the gauged watersheds of the Mendoza River Basin, the drainage area of Reach 84 contains the largest area covered by glaciers. Subbasin 90 is a single gauged small subbasin, so it is a good example of the streamflow response to melt parameter distribution. There was a considerable improvement in the model performance when using the modified model instead of the lumped model in Reaches 79, 84, and 90.
Figure 4 shows that the peaks simulated by the modified model match the observed peak flows. This level of accuracy is only achievable by adjusting the melt parameters for seasonal snow and permanent snow for each elevation band since the lumped model was not able to capture the peaks by calibrating the model for many different sets of melt parameters. Although the results from the modified model show significant improvement in highly glaciated areas and negligible improvement in less glaciated areas, the simulated monthly flow at the main outlet (Reach 86) does not show any improvement with the modified model. It can be seen in
Table 8 that the model performance in simulation of downstream flow (e.g., Outlet 86 in the Mendoza River basin) declines when applying the modified snow algorithm. The new approach enabled calibration of the model versus the flow from upland catchments and optimization of the simulation results by adjusting the parameters separately for subbasin/elevation bands, while in the lump method the parameters were adjusted in order to get the reasonable result and were not necessarily the best at the gauged streams. It seems that the parameter value optimization for upland basins negatively affects the downstream flow in comparison to the lump parameter adjustment method. In other words, factors other than snowmelt could affect the downstream flow, and further model setup and parameter adjustment (groundwater, plant growth and soil moisture parameters, etc.) are necessary for lowlands and flat areas, which were out of the scope of this study. Another reason for declining the accuracy of downstream flow simulation could be the use of inaccurate reanalysis precipitation data. While snowfall/melt is the dominant hydrologic process in dry highlands (highland areas of Mendoza are very dry), flow is less affected by precipitation events, and snow parameter adjustment improves the flow simulation considerably. In lowlands, where the rainfall-runoff model is dominant, the flow is directly under the influence of precipitation events and any problem in input precipitation data could directly decline the accuracy of downstream flow simulation. So, when the results for snowy basins show improvement, flow in lowland areas may not necessarily improve. Comparisons between observed and simulated monthly streamflow during the validation period indicate range of model performances from good to unsatisfactory.
Comparisons between observed and simulated monthly flow from Chilean river basins during the calibration period indicate poor to satisfactory simulation. The RSR and NSE values are generally lower than those values obtained in simulation of monthly flows from the other river basins in this study. Winter precipitation (October to March) inter-annual variability in the Andes is linked to ENSO (El Niño–Southern Oscillation) events and consequently reveals a more complex response of streamflow [
47], which may explain poor model performance in simulation of seasonal flow variability in the Mendoza River Basin and the Chilean river basins in the dry central Andes.
4. Conclusions
Treating the glaciered and un-glaciered areas in the watersheds separately improved SWAT model performance significantly in simulation of volume and variation of runoff in glaciered areas. Spatial and temporal variations of melt rates mainly depend on the spatial and temporal variations of melt factors in hydro-glacial models. While temporal variations of melt factors have been considered in the SWAT model in the past, there has been no consideration of spatial variations in melt factors and lag time factor, which are directly influenced by surface type (snow and ice). In this study, these spatial variations were specifically taken into account.
Model performance using the snow algorithms also depends on the climate of a river basin. Significance of melt water may be negligible in watersheds where the melt season coincides with monsoon precipitation, resulting in no significant difference in the simulation results from the two melt algorithms when monsoons were a factor. For the river basins in the central Andes, applying the modified distributed snow algorithm considerably improved the model performance in simulation of runoff volume in comparison with the lumped snow algorithm, while variation of runoff did not show any improvement. This may be due to high inter-annual and annual variability of flow in these regions, which are more dominated by rainfall-runoff relationships than snowmelt. We can conclude that considering the spatial variations of associated melt parameters significantly improves the SWAT performance in simulation of runoff volume and its seasonal variation in highly glaciated river basins.
While distributed process-based energy budget models have been tested in the SWAT model, no studies have been done to incorporate enhanced temperature-index models into SWAT. Incorporation of an enhanced temperature-index model into the modified snow algorithm of SWAT should be an objective of future studies on enhancing the SWAT snow hydrologic process. The modified snow algorithm did not enhance SWAT model performance in simulation of the streamflow at the main outlet of the river basins. Based on this, one might be tempted to draw the conclusion that at the macro scale the lumped model is preferable to the distributed model. This, however, requires further investigation into modeling such hydrologic processes as glacier surface mass balance and transient snow line altitude. Therefore, it is suggested that the model not only be calibrated using runoff but also be calibrated against the individual snow hydrology components.