Streamflow evaluation using IHACRES model in Kelantan river basin, Malaysia

Kelantan is a flood-prone area where in the past years flood had occurred quite frequently. Determining a hydrological model that can represent Kelantan River basin by giving plausible simulated runoff according to the observed runoff is essential as this will allow appropriate prediction of future flood by using forecasted rainfall and other data. In this study, the IHACRES model was used to simulate runoff and the calibrated simulated runoff by daily scale and seasonal flood events were compared with observed runoff. In general, the IHACRES model performed better in seasonal scale as compared with annual scale in terms of calibration. However, performance of IHACRES degraded during validation stage, whereby the model tends to underestimate the high peak flows but estimate rather more accurate when no peak flows were present. In terms of annual scale, the best model was obtained by calibrating the streamflow in 2012 – 2013 (2 years), the validation results were not satisfactory with NSE = 0.473 and PBIAS = 27.7%. On the other hand, for seasonal analysis, the best model was obtained by calibrating the data of NEM 6 (November 2017 – March 2018). 3 out of 5 of the validation periods show unsatisfactory results (NSE ≤ 0.50). NEM 1 (November 2012 – March 2013) show the best validation results with NSE = 0.853. Further calibration is required in order to enhance the accuracy of the model.


Introduction
Kelantan, which is located on the east coast of Peninsular Malaysia, is historically a floodprone area especially during Northeast monsoon that causes great inconveniences to the locals. For example, the monsoon flood that occurred in 2014 [1] is described as the worst flood in decades. Flood happens due to the large volume of water that exceeds the channel capacity and extreme events such as prolonged heavy rainfall. Proper management of water and environmental resources would be paramount to tackle the problem of floods.
Hydrological modelling or rainfall-runoff (RR) model has become an important tool to allow the understanding of different hydrological processes and the prediction of the behaviour in a real-world system [2]. Apart from that, they are also paramount tools that are useful for a variety of applications, e.g. flood prediction, flood design, water resources planning, etc. [3]. Simulation of streamflow with RR models have been used extensively all over the world [2,[4][5][6]. According to Devia et al. [2], RR models are currently required and paramount to act as a tool for the management of water and environmental resources. Development of RR modelling is also needed to forecast the hydrological output such as the runoff hydrograph of an area [7]. Also, RR modelling has even gained its importance in urban areas with Salvadore et al. [8] emphasizing the highly challenging in urban catchment modelling and the rapid growth of urban population and area.
As flooding events are increasing in terms of frequency and impact, it is critical to there is a need to determine a RR model (with the most recent data) that can represent Kelantan River basin by giving simulated runoff that is plausible according to the observed runoff. A plausible RR model will allow the appropriate prediction of future flood by using forecasted rainfall and other data. The model parameters that were determined from the simulation can also act as a reference for other studies that look into the characteristics of Kelantan River basin such as soil properties, etc.
In this paper, the metric conceptual IHACRES model is applied to the Kelantan River basin. The IHACRES is a lumped conceptual model, which simulates rainfall-runoff response of catchments to total stream flow, with calibrated parameters prior to simulation by comparison with observed stream flow data [9]. This model had been adapted in catchments of Australia, South Africa, Jordan, and Thailand and satisfactory results of modelled streamflow were presented [5,[10][11][12]. For present study, the rain gauge data and streamflow data obtained from the Department of Irrigation and Drainage (DID), Malaysia and temperature data from the Malaysian Meteorological Department (MMD), starting from year 2012 to 2018 were utilized in the IHACRES model to evaluate the streamflow on the basis of daily scales as well as seasonal events.

Study area and data acquisition
Kelantan is one of the fourteen states in Malaysia and is located in the north-eastern corner of Peninsular Malaysia with coordinate of latitudes 4° 40' to 6° 12' N, and longitudes 101° 20' to 102° 20' E. The longest river in Kelantan is the Kelantan River spanning 248 km long and drains an area of 13,100 km 2 , which occupies more than 85% of the whole state. The maximum length and breadth of catchment are 150 km and 140 km respectively. There is about 95% of the catchment consisting of steep mountainous country rising to a height of 2,135 m, and the rest is undulating land. The mountain ranges in the eastern and western portions have granitic soil cover with mixture of fine to coarse sand and clay. Whereas in the extreme east and west of the southern half of the basin, fine sandy loam soil can be found. The remainder of the area has various soil cover with depth ranging from few metres to more than 9 metres [13].
Kelantan is prone to flood for almost every year especially during November and December. According to Alias et al. [14], rainfall is abundant in the east-coast during the north-east monsoon, which causes most of the major historical flood events. Situated in the east-coast of Peninsular Malaysia, Kelantan had faced several historical flood events in the past. In December 2014, flood known as the Kelantan Big Yellow Flood 2014 occurred in Kelantan with flood levels reaching 5 to 10 meters. A flood known as the worst natural flood caused by monsoon rains hit Kota Bharu, the capital of Kelantan in November 2005 [15]. Other flood events also occurred in 1994 and so on that caused damages and death. In Kelantan River basin, there are 54 operating rain gauge stations, 6 streamflow stations, 3 weather stations (which measure temperature, relative humidity, and wind speed), and 2 evaporation stations in Kelantan river basin. The stations are depicted on Kelantan map in Figure 1. For the present study, only the daily rainfall, streamflow and temperature data starting from 1st January 2012 until 31st December 2018 (7 years) were acquired.

Model description and structure
IHACRES (Identification of unit Hydrographs and Component flows from Rainfall, Evaporation and Streamflow data) is a lumped catchment-scale rainfall-streamflow modelling tool that can predict streamflow and depict the dynamic relationship between rainfall and streamflow by using rainfall, temperature (or evaporation), and streamflow data. The time step of the data can be either in minutes, hours, or days. The present study employs the IHACRES Version 2.1.2 or the IHACRES Classic Plus, which is a redevelopment of the first original version of IHACRES which was developed in 1994 [16]. According to Croke et al. [17], this version retains all original features with some additional features, including the extension for ephemeral catchments, extension of linear routing module, cross correlation tool, additional goodness of fit indicators, and visualisation tools. In this study, the daily rainfall, temperature, and streamflow data starting from 2012 to 2018 with units of millimetres, degrees Celsius, and cubic metres per second (cumecs) respectively, were input into the model. As IHACRES is a lumped model, the value of each input data for each day is taken as the average of all the available stations measuring the required data. For instance, the input value of rainfall for a particular day is the average of all 54 rain gauges on that day. Similar method is applied for temperature data. For streamflow data, the data collected at the station labelled as KS4 as shown in Figure 1 is used. This station is situated at the northern part of the Kelantan River system. The Kelantan River system originates from the southern area such as Tahan mountain range for Lebir River and central mountain range for Nenggiri River [13]. Then the river flows northward to the Kelantan River and finally discharge into the South China Sea in the north.
According to Littlewood et al. [18], IHACRES has been applied in catchment with sizes as small as 490 m 2 to 10,000 km 2 . Since the input unit of streamflow in this study is cumecs, catchment area is required as an input in the model. The catchment area for Kelantan River basin at the observation point of streamflow (at station KS4) is estimated by using ArcMap 10.5 software and has a value of 10934.95 km 2 .
The model structure of IHACRES consists of the non-linear module and linear module and is shown in Figure 2. The non-linear module requires input of rainfall and temperature data and convert them into effective rainfall. Then, the linear module transforms the effective rainfall into modelled streamflow by using any configuration of stores in parallel and/or in series [19]. The stores configuration can either be one store only that represent ephemeral streams, or two stores in parallel that allow the representation of base flow/slow flow and quick flow. According to Jakeman and Hornberger [9], configuration of stores other than the aforementioned rarely improves the goodness of fit of the modelled discharge.

Performance indicator
In this study, the Nash Sutcliffe Efficiency (NSE) and percent bias (PBIAS) are used as the performance indicator for the output of IHACRES. According to Moriasi et al. [21], the study evaluated the data from numerous hydrological models and recommended the statistical performance measure for several evaluation criteria. The recommended statistical performance measure as shown in Table 1 is used as a guideline for the evaluation of goodness of fit of the IHACRES model in this study.
NSE is the measure of fit between observed streamflow and modelled streamflow. NSE has value ranging from -∞ to 1. When NSE = 1 the simulated streamflow is deemed as a perfect match to the observed streamflow. NSE = 0 means that the prediction by the model is as accurate as the mean of the observed streamflow, whereas NSE < 0 indicates the mean of the observed streamflow can predict better than the model. In other words, the model prediction is more accurate as the NSE approaches 1.
Another indicator is the relative bias or percent bias (PBIAS). Relative bias is based on the bias indicator. In IHACRES, bias is the measurement of the overall error in the flow volume in mm per year. Whereas PBIAS indicates the average tendency of the simulated streamflow to be larger or smaller than the observed streamflow. PBIAS has the unit of percentage (%), where the optimal value is 0%. The value of PBIAS can be either positive or negative which indicates overestimation or underestimation respectively. The lower the magnitude of the value, the more accurate the model simulation is.

Overall simulation evaluation at daily scale
Evaluation of IHACRES model in daily scale was carried out by calibrating several years of daily datasets and validate the calibrated parameters with the remaining years of data. According to Ye et al. [12], usually 2 years of data are calibrated for IHACRES model. In this study, the 7 years of data were divided into different combination of consecutive 2 years period. This is to allow the parameters to be exposed to inter-annual variability [11]. In the study by Dye and Croke [10], 5 years period of datasets were calibrated by using around 1 year of data and the remaining years were used for validation.
The calibration of the IHACRES involved the linear module and non-linear module. In this study, the parameters of the linear module are estimated by using IV estimator in the model whereas the 5 non-linear module parameters were calibrated with "Grid Search" function in the model. Cross-correlation function in the model shows 1-day delay between observed rainfall and streamflow response. Figure 3 shows the time series streamflow simulation with the best model fit statistics  Referring to Figure 3, the simulated streamflow from the model seems to fairly match the observed streamflow during the calibration period of 2012 to 2013, the values of NSE and PBIAS during this calibration period fall into the "good" range of value for the recommended statistical performance measure of hydrological model by Moriasi et al. [21] ( Table 3). At the start of 2012, the model underestimated the simulated streamflow before gradually matching the observed streamflow after about two months. Also, among the three significant peak flows on January, March, and December of 2013, the model underestimated the flows for March and December. According to Dye and Croke [10], the linear module of the model requires some times in order to build up the slow flow component to an adequate level.
During validation stage, the results were not satisfactory with NSE = 0.473 and PBIAS = 27.7%. It seems that the IHACRES model performs rather poorly when significant peak flows are present. This is in contrast with the study by Ye et al. [12] that concluded the degradation of results from calibration to validation is less for IHACRES. According to Chiew et al. [22], their study found extremely poor daily low flow simulation due to the difficulty in calibrating the parameters involved in "slow flow" component of runoff. In a study by Post et al. [23], slow flow component was removed from the IHACRES model structure and IHACRES performed reasonably well. In this study, the "slow flow" component is part of the linear module that was calibrated only by using IV estimator, hence possibly contributing to the poor simulation performance. Apart from that, referring to Figure 3, the highest peak event was noted to be happened around December 2014 to January 2015, which is also the largest flood event in the history of Kelantan was reported [1]. In our opinion, the lumped model may have limitations in simulating the streamflow at larger scale. Further calibration of the model may be required in this case.

Seasonal flood evaluation
In the east coast of Malaysia, where Kelantan river basin is situated in, rainfall was more abundant and intense during Northeast monsoon (NEM) [14,24]. Catastrophic flood event in Kelantan such as the flood on 1976 and 2014 occurred also during NEM [14,25]. Thus, in present study, evaluation of the IHACRES model was also done based on seasonal scale, specifically focusing on NEM. For this section, single event-based dataset was utilized for E3S Web of Conferences 347, ICCEE 2022 0 (2022) https://doi.org/10.1051/e3sconf/202234704008 4008 calibration and validation [5]. The study period (2012 -2018) in present study covers six NEMs, as shown in Table 5. The calibration process was carried out by using one of the NEMs, and then the rest of the NEMs were used as validation. Each NEM was calibrated until the best results among all sets of calibration and validation were obtained. Calibration of IHACRES by using NEM data might produce results that can reflect the characteristics of the Kelantan River basin in producing streamflow during flood-prone periods.  Table 3 and Fig. 4 show the best set of calibration and validation results obtained by calibrating the data of NEM 6. This combination is the only set of calibration and validation that produced all positive values of NSE. The calibration by using other NEM would either produced validation results of negative NSE values or unsatisfactory NSE and PBIAS values. From the calibration period of NEM 6, the simulated streamflow graph fits fairly well with the observed streamflow with NSE = 0.809 and PBIAS = -2.8% which represent a very good performance of the model. Generally, for the validation part in seasonal scale, the IHACRES model seems to perform rather poor in predicting streamflow whenever surges of streamflow occurred in Kelantan River basin. This is in agreement with the study by Hassan et al. [26] where they found IHACRES model to simulate flows rather poorly when peak streamflow occurred. NEM 1 shows the best result among 5 validations, whereby the simulated streamflow seems to fit almost perfectly with the observed streamflow, with NSE = 0.853 and PBIAS = 13.9%. The slightly higher PBIAS may be due to the surges of streamflow occurred in around January and March 2013. Similar condition also occurred on NEM 2. Even though satisfactory performance was obtained (NSE = 0.527) from this event, but the high peaks of streamflow in around December 2013 and mid-January 2014 were not well simulated by the model. NEM 3, one of period with the poorest NSE results, is the period where the historical Kelantan big yellow flood occurred. The highest streamflow surge during this flood event was significantly underestimated by the model. Next, for NEM 4 and NEM 5, the periods where the lowest and highest average seasonal rainfall were observed respectively, the IHACRES model seems barely able to simulate streamflow that match with the observed data in overall, not to mention of the streamflow peaks.

Conclusion and Recommendations
The use of IHACRES model in Kelantan river basin is generally satisfactory in terms of calibration. Calibration in annual scale and seasonal scale both produced good and very good model performances respectively. However, during validation, performance of IHACRES degraded where most validation results fall to the range of "unsatisfactory" in terms of NSE and PBIAS values. Based on the comparison between simulated and observed streamflow, it is found out that IHACRES did not perform outstandingly in predicting flows especially when high peak flows were present. IHACRES tends to underestimate high peak flows but estimate rather more accurate when no peak flows were present. Since Chiew et al. [22] discussed the extremely poor daily low flow simulation of IHACRES due to the difficulty in optimising parameters involved in "slow flow" component, it is also possible that the poor performance of IHACRES in simulation for this study is due to the linear module calibration method (IV estimator).
In terms of the hydrological model analysis, despite of the lesser input data needed, and lesser parameters needed to be calibrated, IHACRES as a simple type of model does not provide reasonably well performance especially in simulation for Kelantan river basin. The "slow flow" component of IHACRES should either be removed from the model structure, such as in the study by Post et al. [23], or to be calibrated with using the Fixed transfer function in the model by obtaining prior information about the parameter values of linear module using regionalisation method [17]. In this study, the catchment size at the streamflow observation station has a value of 10,934.95 km 2 , which is slightly larger than previously applied catchment sizes that range from 490 m 2 to 10,000 km 2 [18]. IHACRES as a lumped model might perform better for catchment with size smaller than the Kelantan river basin, and perhaps a different hydrological model that is personalized for large catchment size (>10,000 km 2 ) should be used or Artificial Neural Network model (ANN) could be used such as in the case of Hassan et al. [26]. This work was supported by the Universiti Tunku Abdul Rahman Research Fund (IPSR/RMC/UTARRF/2020-C2/C04).