Assessing the Effect of Watershed Spatial Characteristics on Regional Calibration for a Single Event Flood Model: The Case of Kohsukhteh Catchment

1. Application of hydrological models is one of the common methods for quantitative
analysis of watersheds. A class of these models is used to simulate rainfall-runoff
processes. A hydrological rainfall-runoff model deals with integration of time series
data, operational parameters, local variables and physical governing system laws in
order to simulate runoff and other processes of the catchment. The flood hydrograph is
an important graphical representation in hydrological analysis. The shape of hydrograph
is a direct response to management strategies basin. In fact, watershed management is
not possible unless hydrological characteristics of the basin could be fully understood.
HEC-HMS model is a rainfall - runoff model with either a simple lump or a quasidistributed
structure. The model was developed by Hydrologic Engineering Center of
US Army Corps of Engineers. One important capability of this model is the possibility
of its integration with ArcGIS. The HEC-GEOHMS extension automatically calculates
most of the geospatial data that needs to be calculated manually. Its user-friendly
interface offers several tools/icons such as the basin, reach, junction, etc. to enter the
data. The component-based set up of the model makes it possible to modify the
structure of the modeling system by the user. Different options for calculating the loss
method, transmission, and streaming flow are available in HEC-HMS model. Each
element has its own parameters or inputs. This flexible system provides an opportunity
to apply the model to the regions with limited access to data. For a similar reason,
models like HEC-HMS are useful tools to investigate the effects of changes in some of
the spatial and temporal variables. In this study, the effects of the spatial scale on the
optimization of HEC-HMS have been investigated.
The modeling parameter sets obtained by large scale setting generally outperform
local scale hydrologic parameterization. Most research in the field of calibration of
hydrological models has focused on the local parameters but if the at larger scales, it
can be regionalized to other places. In such procedures, the sensitivity analysis is initially used to identify the most important parameters. The aim of this study was to
compare local and large scale calibration.
2. Material and Kohsukhteh watershed is geographically located between 50 °40' to 51 ° 20 ' East
longitude and 31° 20' North and 32 ° latitude. The study area covers an area of 2783
square kilometers. The minimum and maximum elevation values in this basin are 1705
and 3398 meters respectively. The average slope of the area is 19%. The mean recorded
annual precipitation (Shahrekord synoptic station) is 320 mm.
Data preparation for HEC-GEOHMS was the first stage of modeling. Different physical
characteristics, base maps and time series data of the basin were introduced to the
model. The model set up was implemented by setting Loss, Transform and Base-flow
methods for each sub-basin. The necessary steps for meteorological model and control
specifications were taken as well. The 6-hour rainfall and 6-hour discharge time series
were primarily used for the analyses. The model was parameterized at both local and
large scale levels. Parameters related to Loss method, Transfer method and Routing
were selected on the basis of model performance. In order to identify the most sensitive
parameters, sensitivity analysis was performed. Given that the accuracy of the input
parameters is an indicator of the efficiency of modeling and the results of the final
simulation, the results depend on this stage. After identifying the sensitive parameters,
they were selected to optimize the model locally and in a large scale. Despite the
uncertainties and errors of the models, no computer model can expect a complete and
accurate prediction. Therefore, they should be calibrated and validated. To evaluate the
performance of models with different conditions, six events were chosen for the
calibration and two for validation of the model.
3. Sensitivity analysis on the parameters of the model showed that Lag time parameter and
initial abstraction had the highest sensitivity. Calibrated values of the model parameters
were obtained for both calibration and validation events. The results of calibration in six
events and the observed and simulated runoff volume in the basin were compared. Two
events were selected to validate the calibrated model. Runoff discharge and volume for
the basin were simulated. The values obtained by simulation were compared against the
recorded observational data at the hydrometric stations were compared. After
calibration and optimization of the model at local scale, the results showed that both
estimated peak discharge and the time to peak flow, was performed better than obtained values for the basin scale simulation.
The inter-comparison among the events also show some differences in model
performance. Event December 25, 2012 showed the highest accuracy and March 30,
2012 was simulated with the least accuracy. The difference may be due to the change in
soil permeability during winter time. In the case of large scale model, the Event
December 25, 2012 was simulated with high efficiency. Lag time, initial abstraction,
curve number and impervious ratio were identified as the most important factors for calibration. The obtained Nash-Sutcliffe for local calibration was 0.85 while it dropped
to 0.65 for large scale calibration.
4. Modeling is an indirect method that is much faster than field methods. In order to
achieve accurate results in modeling, we need to estimate the model parameters as well
as the time and place of variables. Calibration refers to the process of comparing of the
measurements against the estimated values. The use of local scale calibration versus
large scale calibration has a higher apparent accuracy, and this is in line with
Samaniego, Kumar and Attinger (2010) and Hundecha and Bárdossy. (2004). However,
the commutated value for Sutcliffe model efficiency is less variable in the case of local
calibration. It was demonstrated that the simulated peak runoff was closer to
observations in local calibration compared to the large scale calibration. A similar was found with the simulated runoff volume using local calibration. Although, both
calibration settings provided an acceptable response to the estimation of runoff; the
obtained parameters at basin scale show less spatial sensitivities. This makes it possible
to generalize the calibration results to nearby areas with limited data. The results show
that the large-scale parameterization actually has less spatial dependency and therefore
may provide more reliable results. This finding is in line with reported advantages of
large-scale parameter estimations, in term of saving the time (Troy, 2008; Pokhrel,
2008; Beven, 1002). Sensitivity analysis also showed that parameters such as
imperviousness and initial abstraction had a high sensitivity in the study area.