Impact assessment of land use/land cover and climate change on streamflow regionalization in an ungauged catchment

Accurate streamflow measurement records are very much necessary for various events of water resources planning and management (Blöschl 2016). One way of getting such streamflow records is to measure the streamflow on a particular gauging site for a long period. However, in most cases, this is not feasible due to a variety of reasons such as financial, logistics or purely due to lack of intent (Blöschl 2016). This leads to the fact that a large portion of the total number of river basins present around the globe are either poorly gauged or completely ungauged (Sivapalan et al. 2003; Young 2006; Blöschl et al. 2013). In addition, the hydrological parameter observation networks are continuously weakening (Mishra & Coulibaly 2009). Furthermore, the number of gauging stations in comparison to the large river basins are very low. Hence, small watersheds are devoid of water resources assessment data (Ibrahim et al. 2015).

Prediction of hydro-climatological variables for a drainage basin requires modelling, where hydrological models simplify the complex hydrological processes occurring within the basin and present them in a simplified manner. Estimation of streamflow in gauged and ungauged catchments is being carried out using either distributed physically-based models, conceptual and semi-distributed models or data-driven models (Razavi & Coulibaly 2012). This is only possible where long-term hydrological records are available, i.e., gauged catchments. In the case of ungauged catchments, regionalization is the method for estimation of different variables like streamflow, sediment load and water quality parameters. The term regionalization can be explained as follows: it is the transfer of an optimized model parameter set from one or more gauged (donor) catchments to an ungauged (receiver) catchment (Blöschl & Sivapalan 1995; Oudin et al. 2010). Regionalization is categorized into two types: hydrologic model-dependent and hydrologic model-independent (Razavi & Coulibaly 2012). Spatial proximity, physical similarity and regression analysis are some of the widely used regionalization techniques across the globe in both categories (Nathan & McMahon 1990; Sefton & Howarth 1998; Yu et al. 2002; Merz & Blöschl 2004; Parajka et al. 2005; Götzinger & Bárdossy 2007; Oudin et al. 2008; Samuel et al. 2011; Zhang et al. 2015; Müller & Thompson 2016; Swain & Patra 2017a, 2017b).

Considerable progress has been observed in the PUB (Prediction in Ungauged Basins) initiative during the last decade. However, most of the studies were case studies conducted in well-gauged catchments situated in well-developed countries (Visessri & McIntyre 2015). On the other hand, very few studies address the issue of the non-stationary environment during regionalization. Specifically, in developing countries, where the land surface is undergoing rapid change, the issue of non-stationarity holds great importance. Land use/land cover change (LULCC) and climate variability are important elements out of the many active environmental factors affecting catchment hydrology (Chen et al. 2012; Zhang et al. 2016). The reason for this may be that land use change affects evapotranspiration (ET), interception and infiltration. Thus, ultimately surface and sub-surface flows are affected (Niraula et al. 2015). Whereas climate variability brings distinctive changes in hydrological regimes altering spatial and temporal patterns of water resources in an area (Khoi & Suetsugi 2014). A few cases dealing with the non-stationarity during regionalization are discussed below.

The process of predicting the hydrologic response and streamflow response to land use change provided good results [Nash-Sutcliffe efficiency (NSE) ≤ 0.85] when Croke et al. (2004) applied the ‘Identification of unit Hydrographs And Component flows from Rainfall, Evaporation and Streamflow data’ (IHACRES) and CATCHCROP model to the ungauged catchments in Thailand using a simple regionalization and scaling technique. Singh et al. (2011) introduced the idea of trading-space-for-time to counter the changing climate during continuous streamflow prediction and was successful in implementing it during the study of 394 US watersheds. The transferability of hydrological model parameters under non-stationary conditions was investigated by Li et al. (2012) using differential split-sample tests and Monte Carlo simulation in combination with Dynamic Water Balance Model and SIMHYD models in 30 catchments in Australia. Transferring model parameter values obtained from dry periods to wet periods resulted in smaller errors in streamflow prediction. Development of a multiple regression model using few catchment characteristics led to efficient estimation of the non-stationary regional flood quantiles [Q₅(t) and Q₁₀₀(t)], with relative root mean square error (RMSEr) of 38.2% and 60.8% respectively during the application of a non-stationary regional model in 29 catchments of Canada and Northern USA by Leclerc & Ouarda (2007). Lima & Lall (2010) applied a hierarchical Bayesian model using the scaling relationship between catchment properties which resulted in a competent estimation of monthly flood flow probability distribution parameters for 44 ungauged sites in Brazil in non-stationary settings. Singh et al. (2014) tested the combination of streamflow signature regionalization with the trading of space for time approach for robust hydrologic projections in the Olifants basin in southern Africa under changing climate condition. The statistically downscaled General Circulation Model against observed climate data led to minor deterioration in reliability of runoff projections for the historical period. Visessri & McIntyre (2015) demonstrated the use of the regression method against catchment properties for regionalization of three streamflow signatures (runoff coefficient, Base Flow Index, and seasonal elasticity of flow) under land use change and variable data quality in 44 sub-catchments in Thailand. The regression equations derived from regionalization failed to predict the streamflow indices except for runoff coefficient under non-stationary conditions. Halder et al. (2015) and Tian et al. (2014) are the two recent studies on land use change in the Indian context but are deprived of streamflow analysis in ungauged catchments. Li et al. (2009) and Yin et al. (2016) used the Soil and Water Assessment Tool (SWAT) for the Heihe catchment and Jinghe River basin, respectively, to analyze the impacts of land use/land cover (LULC) and climate variability on streamflow, producing satisfactory results. Even though the non-stationary issue was well addressed in the above two cases, the catchments are gauged. Hence, very few documents are available at the moment related to streamflow estimation in ungauged catchments considering the non-stationary environment.

The importance of streamflow estimation in ungauged catchments across the globe is a well-documented subject. However, in a developing country like India, the issue of continuous streamflow regionalization is not prevalent. The additional factors like changing climate and LULC make the regionalization studies more challenging. Considering the above situation, the objective of this study is to enumerate the isolated as well as combined effects of LULCC and climate variability on streamflow and associated uncertainty at an annual scale in ungauged catchments using hydrological model-dependent regionalization techniques. The hydrological model in question is SWAT, while the two regionalization techniques are Inverse Distance Weighted (IDW) and Kriging. Application of SWAT for quantifying the non-stationary effect on prediction is acknowledged in previous studies by researchers. However, application of this procedure on ungauged catchments is a relatively new subject which needs further attention.

India is home to a diversity of climatic regions. Thirty-two catchments (Figure 1) from Eastern and Southern India, spreading over nine river basins, were selected as the study area. All these catchments are headwater regions of nine different river basins, where flow regulation is minimal. The selected 32 catchments are spread over the following states: Jharkhand, West Bengal, Chhattisgarh, Odisha, Telangana, Andhra Pradesh, Tamil Nadu, Karnataka, and Maharashtra. Mean annual rainfall in the first four states is 1,400 mm and for the rest, it is around 1,000–1,100 mm (http://www.imdpune.gov.in/). The key reason for rainfall in India is the southwest monsoon, from July to October. Apart from that, the southern part of India receives a good share of annual rainfall during October and November due to the northeast monsoon. The central and eastern parts of India experience a high temperature of 44 °C during May and June, whereas December and January are the coldest months with temperatures as low as 5 °C. In southern India, summer is hot and humid with a high temperature of 45 °C. However, winter is pleasant with a minimum temperature of 15 °C. The topography and geographical area of the catchments in these regions accommodate large variation.

Daily rainfall data of the study area were collected from India Meteorological Department (IMD) for the period of the year 1990 to 2011. Only those rain gauge stations where a maximum of 25% data were missing were considered. Similarly, daily streamflow measurement records were collected for the same period from the India-WRIS database (http://www.india-wris.nrsc.gov.in/wris.html) which is jointly operated by Central Water Commission, India (CWC) and Indian space research organization. The land use database was collected from the Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC) (https://daac.ornl.gov) and soil data were collected from the Food and Agriculture Organization of the United Nations (http://www.fao.org/soils-portal/).

The need for calibration of a hydrological model is to determine the optimized parameter set which might have an impact on the catchment response. These parameters may or may not have any physical meaning, but they have a high level of uncertainty associated with them (Heuvelmans et al. 2004). The first period (1990–2000) was considered as the calibration period, while the second period (2001–2011) was treated as the validation period. Before proceeding with the calibration, sensitivity analysis was performed to identify the most sensitive parameters out of a group of 25 model parameters. The parameters used for calibration are presented in Table 1 with the physical significance of each parameter and range of parameters used. SWAT-CUP was used to calibrate the model against observed discharge. In the present study, the Sequential Uncertainty Fitting (SUFI-2) tool was used to assess the uncertainty associated with the prediction. In SUFI-2, uncertainty data associated with input parameters are represented as uniform distributions, however model output uncertainty is quantified by the 95% prediction uncertainty (95PPU) calculated at the 2.5% and 97.5% levels of the cumulative distribution of output variables attained through Latin Hypercube sampling (Abbaspour et al. 2007; Yang et al. 2008). SUFI-2 represents uncertainties of all sources through parameter uncertainty in the hydrological model. It is a bracket of good solutions represented by the 95PPU, generated by certain parameter ranges. P-factor and R-factor are the two statistical terms used for quantification of the uncertainty in terms of a single signal. P-factor is the percentage of observed data bracketed by the 95PPU and R-factor is the average thickness of the 95PPU band. Theoretically, the value for the P-factor ranges between 0 and 100%, whereas R-factor ranges between 0 and infinity. For P-factor, a value of greater than 70% is recommended, whereas having an R-factor of about 1 for discharge calibration and validation is good (Abbaspour 2015). 95PPU should envelope most of the observations, at the same time it is desirable to have a small envelope.

The distance between the two centroids of the gauged and ungauged catchment is represented by ⁠.

For estimation of model parameters in ungauged catchments, all of the 32 catchments were treated as ungauged, one by one.

Agricultural intensification, urbanization and deforestation in addition to other human interventions are some of the primary reasons for land use change. The impact assessment of land cover change on streamflow regionalization was analyzed using land use decadal data of India for the years 1995 and 2005 (Roy et al. 2015). Figure 2 shows the two LULC maps of India for the years 1995 and 2005. The supervised image classification approach with the maximum likelihood classification tool was implemented to classify the LULC image into 19 classes. Forest and agricultural lands are the two major land use classes for most of the catchments under study.

The impact of LULCC on streamflow was analyzed by generating different cases mentioned in Table 2 for calibration and validation periods. SWAT simulations were carried out using land use data of the year 1995 for period 1 and land use data of the year 2005 for second period. The two selected land use maps were able to represent the scenario of the respective periods quite effectively. The cases mentioned above may be explained as follows: case 1 represents a scenario where simulation was carried out using LULC data of the period 1 along with the climatological data of the same period. However, in the second case, simulation was performed using climate data of period 1, and land use data of period 2. According to Li et al. (2009), during the simulation, one factor is variable while others are kept constant, which explains the effect of the variable factor on runoff simulation. Case 1 represents the ‘base’ scenario where the impact of LULC change along with climatological factors upon runoff can be quantified. The difference between case 1 and 2 will quantify the isolated impact of LULC change on streamflow. Case 3 uses land use data of the first period while the climatological data are from the second period. Nonetheless, case 4 was performed using both climatological and land use data of the second period. As explained above, the contrast between different cases quantifies the effects of the above-mentioned two factors on streamflow. However, the four cases mentioned were repeated for the catchment when it was treated as ungauged. The percentage change in streamflow due to LULC and climate change was estimated on a yearly scale.

Based on a previously published recommendation (Moriasi et al. 2007), NSE, RMSE-observations standard deviation ratio (RSR) and Percent Bias (PBIAS) were chosen as indicators to judge the model performance.

The change in LULC may be attributed to the alteration which occurred in forest cover and cropland as most of the catchments out of the 32, experienced major changes in the above-mentioned two land use classes (Figure 2). Only two catchments located in the southern part of the study area experienced a major change in built-up area. The extent of change in forest cover for some catchments is significant while it is marginal for others. However, there are few catchments where the change is minimal which can be neglected. One catchment that is highlighted in Figure 1, where the land use change is significant was selected for the presentation of the non-stationarity analysis for simplified depiction of results. Forest cover (41.53%) and cropland (40.58%) are the two dominant components of the land cover for the selected catchment. Forest cover and cropland area decreased (2.26% and 12.31%, respectively) from the first period to the second period. In addition, there was an increase in wastelands by 54.02% during the period of study. Social development and population growth might be the factors contributing largely to the change in land use pattern (Yin et al. 2016).

The Mann–Kendall test explained the climatic variability of the catchment for the selected period (1990–2011). The trend analysis indicates a decrease in annual rainfall volume for two of the rain gauges (station 1 and 4), while no significant trends are observed for the other two (Figure 3). However, according to the Kendall slope, mean annual temperature tended to increase for three weather stations during that period (Figure 4). The overall results suggest that the climate in and around the catchment is getting warmer and drier. The specific results of the Mann–Kendall test are presented in Table 3 which suggests that the decreasing trend in annual rainfall is significant for two stations (station 1 and 4).

The performance of SWAT for the 32 catchments can be termed as satisfactory during calibration and validation periods as per the model performance criteria set by Moriasi et al. (2007). The minimum, maximum, mean and median values of three model performance judging criteria are listed in Table 4. SWAT largely performed better for catchments present on the northern side of the study area than those catchments located on the southern side. The missing records of hydro-climatological data for a few southern catchments might have resulted in such output. Although SWAT performance deteriorated marginally when catchments were treated as ungauged in turn with the application of IDW and Kriging, it is still within the acceptable limit. As mentioned earlier, the effect of non-stationarity due to changing climate and LULC was demonstrated for a single catchment as the representation of the whole analysis. The results of the modeling for this catchment can be put within the bracket of ‘good’ results. As demonstrated below in Figure 5, the simulated monthly streamflow matched with the observed streamflow except for a few peak flow sections. Overestimation of peak flow is observed. The values of NSE, RSR and PBIAS were 0.75, 0.29 and −15.5% respectively during calibration. While during the validation period these were 0.7, 0.33 and −15% respectively. A P-factor of 0.83 and R-factor of 1.06 indicate the output uncertainty is within the acceptable limits. The values of NSE, RSR and PBIAS were 0.71, 0.33 and −17.2% (calibration), and 0.69, 0.33 and −17.55% (validation), when IDW was applied. The Kriging technique produced similar results as the IDW approach with values of NSE, RSR and PBIAS at 0.72, 0.29, and −16.41% (calibration), and 0.71, 0.3 and −16.13% (validation). P-factor and R-factor for IDW and Kriging techniques are found to be above the threshold value that can be termed as ‘acceptable’.

The impact of LULCC and climate variability was analyzed through the four generated cases mentioned earlier. The isolated impact of land use change was estimated by comparing case 1 with case 2. The difference between these two cases indicates an increase in annual runoff of 3.52%. The decrease in forest cover might have played a key role in the increased runoff. The isolated impact of the variable climate was analyzed with comparison between case 1 and 3. A decrease in annual runoff volume of 21.7% is detected. Overall results show that the combined effect of LULCC and climate variability results in a decrease in annual runoff volume of 22.53%. The contribution of climate variability to the change in streamflow overshadows the influence of LULC change. The percentage change in annual streamflow volume due to the combined effect of climate variability and LULC change is greater than that of the summation of the isolated impacts of these two factors. The results indicate that apart from these two factors mentioned above few other hydro-climatological or geomorphological factors might have played a role in the change in streamflow quantity.

The change in simulated streamflow between the two periods was analyzed at an annual scale. IDW and Kriging were applied to transfer the optimized model parameters from nearby gauged catchments to the selected catchment treated as ungauged. An increase in annual runoff volume was observed during IDW and Kriging application due to the isolated effect of LULC change (case 1–case 2). For IDW and Kriging, streamflow increased by 2.56% and 2.6%, respectively, due to the isolated impact of LULCC. However, a decrease in annual runoff volume was observed during application of both regionalization techniques due to changing climate (case 1–case 3). The percentage decreases in annual runoff volume are 21.09% and 21.6% for IDW and Kriging, respectively. The combined effect of these two factors resulted in a decrease in annual runoff volume of 21.8% and 22% for IDW and Kriging, respectively. The results obtained for the ungauged catchment scenario matched closely with that of the gauged scenario.

This study is an attempt to address the non-stationarity issue through change in LULC and climate during spatial regionalization of continuous streamflow in an ungauged catchment. The small change in annual streamflow due to LULC change indicates that human intervention in the catchment area is minimal. The conversion of cropland to fallow land in addition to decreased forest cover might have played an important role in the increase of annual runoff. However, the variation in climate had a prominent contribution towards change in streamflow quantity. Although there was an increase in annual rainfall during the period (1993–1995), a decreasing trend was observed afterward. The variability in streamflow during the study period is more related to change in rainfall during the monsoon period: one reason for this might be the variable rainfall in monsoon-dominated areas is more intense and distinct than LULCC. Even though climate variability has greater impact on streamflow in the catchment, the degree of correlation between rainfall and streamflow was non-linear due to the combined effects. When the catchment was treated as ungauged, IDW and Kriging produced more or less similar results to that of the gauged scenario. Moving from the gauged to ungauged case, a decreased model performance is observed which is known as a spatial loss in model accuracy. Likewise, moving from calibration to validation stage, the decrease in model performance is known as a temporal loss in model accuracy (Parajka et al. 2005). The satisfactory performance of IDW and Kriging indicates that the presence of well-gauged catchments in the proximity of an ungauged catchment is useful.

Regional climate, soil types and vegetation are certain factors that make the regionalization case specific. Most of the SWAT parameters are related to the factors mentioned above. The model concerned and the objective of the study decide the suitability of the regionalization technique. A change in model or parameters mentioned above might have produced different results. It is essential to recognize that there is no single signal representing model output, but rather an envelope of good solutions expressed by the 95PPU, produced by certain parameter ranges. Quantifying the effect of different uncertainties on model outputs separately is highly desired. However, it is very hard to achieve. Hence, the combined effect is enumerated on model outputs. According to the overall degree of prediction uncertainty, the Kriging approach produced marginally better results than the IDW method for this study.

As mentioned earlier, most of the regionalization studies were carried out considering the peripheral atmosphere to be stationary. The result of this study suggests that the ever-changing climate and LULC has an impact on streamflow. Hence, it is highly recommended for hydrologists, engineers and policymakers to consider these changing parameters during planning and management of water resources and ecology, with the associated socioeconomic impacts.

The application of SWAT for streamflow simulation in non-stationary environments is justified by the satisfactory results obtained in terms of NSE, RSR and PBIAS. The performance of the above-mentioned model is equally good when the catchments are treated as ungauged. IDW and Kriging are found to be competitive with each other, as both approaches produced more or less similar results. The presence of well-gauged catchments around the pseudo ungauged catchment is crucial for better performance of these two geographical distance based regionalization techniques. Both LULC and climate varied during the study period constantly. The combined effect of these two factors resulted in a decrease in annual runoff of 22.53%. However, an isolated impact of LULC brought about a moderate increase in annual runoff volume of 3.52%. On the other hand, a decrease of 21.7% in annual runoff volume was observed due to the isolated effect of climate change as the climate is getting warmer and drier. Out of these two factors, variable climate played a predominant role in surface water hydrology. Hence regionalization studies must consider the non-stationary effects during any hydrological predictions for better water resources planning. A comparison between a stationary and non-stationary condition may further highlight the importance of LULCC and climate change studies. Consideration of a few more geomorphological factors in the analysis should be a future objective to identify the impact of other non-stationary parameters on streamflow.

The authors would like to acknowledge the IMD and CWC, India for providing the hydro-climatological data required for this study.