Assessment of Variable Source Area Hydrological Models in the Upper Cauvery Basin, Karnataka, India

Abstract

With increased availability of spatial data-sets of catchment characteristics and hydrometeorological variables, distributed hydrological models are being applied to solve a variety of problems related to catchment hydrology and water resources management. However, experimental results obtained in recent decades have shown the possiblity of existence of runoff generation mechanisms other than the conventional infiltration-excess (Hortonian) mechanism. In particular, it has been shown that Variable Source Area (VSA) mechanism of runoff generation may prevail in humid steeply sloping and well vegetated watersheds. Accordingly, efforts have been made by previous researchers to incorporate this mechanism into distributed hydrological models and their performances have been evaluated in mostly humid temperate regions and not so much in humid tropical regions. The primary objective of the present study was to compare the performances of hydrological models which incorporate the Variable Source Area (VSA) mechanism of runoff generation with that of the Soil and Water Assessment Tool (SWAT) which employs the conventional infiltration-excess mechanism of runoff generation. One of the VSA based model used, SWAT-VSA, has been proposed by earlier researchers as a re-conceptualization of the SWAT model and uses a topography-based wetness index to identify source areas and simulates runoff in a manner consistent with VSA hydrology. In the present study, the topography-based wetness index was replaced with a Modified Normalized Difference Water Index (MNDWI) derived from satellite imagery resulting in a new VSA model version, SWAT-MNDWI. Performance evaluation of the models was carried out through their application in two humid tropical watersheds (Hemavathi – 2974 km2; Harangi – 538.8 km2) located in the Upper Cauvery River Basin (36,682 km2), India wherein previous studies have shown the existence of VSA hydrology. The other aspects addressed in this study include: assessment of significance and magnitude of trends in historical records of observed hydrometeorological variables in the Upper Cauvery Basin, evaluation of uncertainties associated with streamflowii predictions of the 3 hydrological models and simulation of the hydrologic impacts of hypothetical land use/land cover (LU/LC) changes in the Hemavathi and Harangi watersheds. The present study examined the significance and magnitude of trends in the monthly rainfall (33 rain gauges), maximum and minimum temperature (6 climate stations) and streamflow at 4 gauge sites in the Upper Cauvery Basin for the historical 30 year period 1981-2010. The statistical parameters - Coefficient of Variation (CV) and percentage departure were calculated for average monthly values separately for 3 decades. The Seasonal-Kendall and Sen’s slope estimator were used to calculate significance and magnitude of trends in rainfall, temperature and streamflow data. Detrended Fluctuation Analysis (DFA) method was used to detect long-term persistence in the time series data. As expected, the CV of rainfall shows a large variation in the month December to March, while the percentage departure also varies during these months for different decades. But there was no significant trend found for all rain gauge stations and sub basins except for the Arkavathi sub basin. For maximum temperature there was not much variation except in the months of May and June at the Hassan climate station. Statistically significant trend was observed in maximum temperature for Chikmagalur and Hassan stations. The CV of minimum temperature shows a large variability from November to March for all climate stations and also a significant increasing trend for Hassan and Bangalore stations, while for Madikeri a decreasing trend was observed with a variation of -0.16 0C/year. There was not much variation found for streamflow except in K M Vadi gauge site and T.Narasipur gauge site which showed a significant decreasing trend of -0.778 m3/s/ year. Long range dependence analysis revealed a weak persistence for both rainfall and streamflow of the basin. Using relevant data inputs pertaining to rainfall, climate, elevation, Land use/Land Cover (LU/LC) and soils, the SWAT, SWAT-VSA and SWAT-MNDWI models were applied separately to both watersheds using a daily time step. Models were calibrated for the historical period 2000-2003 and validated for the period 2004-2006 using observed daily streamflow records at the watershed outlets. The comparative assessment focused specifically on the following aspects for the six cases considerediii (3 models applied to 2 watersheds): 1) sensitivity of model parameters 2) accuracy of daily streamflow predictions at the watershed outlets 3) predictions of spatially and temporally averaged annual water balance components 4) differences in spatial patterns of source areas of surface runoff. Sensitivity analysis indicated that for the SWAT model, Curve Number (CN) was the most important parameter while for the VSA based models, parameters related to the unsaturated zone and shallow groundwater were important, a result consistent with the runoff mechanism incorporated in the models. The accuracies of streamflow prediction as determined from scatter plots and model performance statistics were more or less similar both in calibration and validation for all the three models with the models performing better in the forested Harangi watershed. Overall, the SWAT-MNDWI model proved to be the best one in simulating daily streamflow with Nash-Sutcliffe efficiency (ENS) of 0.85, coefficient of determination (R2) of 0.88, percentage bias (PBIAS) of 13.2% and root mean square error (RMSE) of 37.48 m3/s for the Hemavathi watershed and corresponding values of 0.88, 0.88, 1.09% and 16.67 m3/s for the Harangi watershed. All three models simulated spatially and temporally averaged major water balance components in a consistent manner resulting in a residual error of <5% of annual rainfall in the annual water balance. However, evapotranspiration loss as a percentage of rainfall appeared unreasonable (27% - 32%) for the wet Harangi watershed probably on account of it being predominantly forested. The spatial patterns of surface runoff generation were somewhat similar for the SWAT-VSA and SWATMNDWI models, but completely different for the SWAT model, again a result consistent with the runoff generation mechanism adopted. Overall results of this study have demonstrated that models incorporating VSA hydrology, and in particular the SWAT-MNDWI model proposed in this study, provide accurate and convenient tools for distributed hydrologic modelling in humid tropical watersheds. This study also focuses on assessing uncertainties associated with SWAT-MNDWI, SWAT-VSA and SWAT models using SWAT-CUP (Calibration and Uncertainty Programs) tool. Two multi-objective uncertainty techniques (Generalized Likelihood Uncertainty Equation (GLUE) and Sequential Uncertainty Fitting algorithm (SUFI-2)) were tested for the Hemavathi and Harangi watersheds. The goodness-of-fit and efficiency of the models have been tested using ENS as the objective function. GLUEiv and SUFI-2 techniques yielded good results in minimizing the differences between observed and simulated streamflows at the outlets of the Hemavathi and Harangi watersheds. The results show that GLUE performance was slightly better than the SUFI-2 technique for all models for both the watersheds during calibration and validation periods. The 95PPU estimated by the GLUE and SUFI-2 techniques are very close to each other and larger than 45% (P-factor) for all models for both the watersheds during calibration and validation periods. For GLUE, R-factor values during the validation phase for the Hemavathi watershed were 0.35, 0.38 and 0.34 for the SWAT-MNDWI, SWAT-VSA and SWAT models respectively with corresponding values for the Harangi watershed being 0.41, 0.39 and 0.40. It should be noted that that both GLUE and SUFI-2 cannot accurately quantify the prediction uncertainty of SWAT-MNDWI, SWAT-VSA and SWAT models. Overall results indicated that the GLUE technique applied on the SWAT-MNDWI model performed best in quantifying the prediction uncertainty of streamflow at the outlets of both watersheds. In order to simulate the hydrologic impacts of LU/LC changes in the study area, two hypothetical LU/LC change scenarios were formulated for Hemavathi and Harangi watersheds. The SWAT-MNDWI, SWAT-VSA and SWAT models were used to simulate the hydrologic responses under these scenarios. Values of average annual water balance components and their percentage change with respect to reference results were calculated for both watersheds using the three models. Additionally, an effort was also made to construct the Flow Duration Curves (FDCs) using daily streamflow values generated under each scenario. Differences in optimal parameters of an empirical model for the FDC, magnitudes of flow quantiles, high flow index and low flow index were computed for each scenario. For the Hemavathi watershed, with increase in agricultural land there is increase in water yield predicted by all three models. With increase in forest cover there is decrease in water yield predicted by SWAT-VSA and SWAT models while for SWAT-MNDWI an increase in water yield was found. For Harangi watershed, with increase in agricultural land or forested area there is decrease in water yield for all three models except SWAT-VSA model in scenario I. Both the scenarios appeared to have significant impacts on the runoffv regime as indicated by significant changes in FDC model parameters, flow quantiles and flow indices. Overall results of this study provide useful inputs with regard the magnitude and direction of likely future changes in important hydrometeorological variables which can be used to prepare plans for mitigation and adaptation to climate change in the Upper Cauvery Basin. The present study has demonstrated an overall methodology for application, performance evaluation and uncertainty analysis of distributed hydrological models using a variety of ground-based inputs and satellite data within a GIS framework. Since previous studies in similar watersheds in the Western Ghats region have identified VSA as a dominant mechanism of runoff generation, the spatial patterns obtained with the SWAT-VSA and SWAT-MNDWI models provide information which will prove to be extremely useful in soil and water conservation measures and in identifying source areas of non-point pollution. The SWAT-MNDWI model proposed in this study is particularly attractive since it employs satellite imagery to accurately identify areas of different wetnesses within the watershed and integrates this information into a distributed hydrological model. As the results of this study have demonstrated, such a modelling approach using VSA hydrology provides an accurate and convenient tool for distributed hydrologic modelling and impact assessment of LU/LC changes in humid tropical watersheds.