Abstract

The River Subansiri, one of the largest tributaries of the Brahmaputra, makes a significant contribution towards the discharge at its confluence with the Brahmaputra. This study aims to investigate an appropriate model to predict the future flow scenario of the river Subansiri. Two models have been developed. The first model is an artificial neural network (ANN) based rainfall-runoff model where rainfall has been considered as the input. The future rainfall of the basin is calculated using a multiple non-linear regression-based statistical downscaling technique. The proposed second model is a hybrid model developed using ANN and the Soil Conservation Service (SCS) curve number (CN) method. In this model, both rainfall and land use/land cover have been incorporated as the inputs. The ANN models were run using time series analysis and the method selected is the non-linear autoregressive model with exogenous inputs. Using Sen's slope values, the future trend of rainfall and runoff over the basin have been analyzed. The results showed that the hybrid model outperformed the simple ANN model. The ANN-SCS-based hybrid model has been run for different land use/land cover scenarios to study the future flow scenario of the River Subansiri.

INTRODUCTION

The River Subansiri is one of the major north bank tributaries of the River Brahmaputra. Originating at the Kangig glacier range in Tibet, it flows through Tibet for 170 km and then 250 km through Arunachal Pradesh before entering Assam at Dulangmukh in Dhemaji district. The river flows for another 130 km through the plains of Assam to join the River Brahmaputra near Jamugurighat in Assam (IRG 2014). The uppermost part of the Subansiri basin is the Tibetan high elevation stretch and approximately 4,000 km2 of the drainage area of the basin falls in an area of perpetual snow (IRG 2014) out of a total catchment area of 34,246.19 km2. Thus, in the upper reaches, the river is fed by snowmelt. The basin experiences the highest temperature from June to the beginning of October and the temperature begins to drop by the end of November (Gogoi et al. 2013). The high temperature during June to October is responsible for melting of snow at the upper catchment. The heavy precipitation that occurs in the Subansiri basin also contributes to the discharge of the river. Due to the northeast as well as southwest monsoon, precipitation occurs in this region in abundant quantities (Sarkar 2015). Thus, in the lower reaches, the river is fed by rainwater.

Climate change is likely to lead to an intensification of the global hydrological cycle and have a major impact on regional water resources (Arnell 1999). Global warming, caused by increased atmospheric concentration of carbon dioxide and other trace gases will alter the radiation balance of the atmosphere. This in turn will cause increases in temperature and changes in precipitation pattern and other climatic variables (Changchun et al. 2008). According to a study carried out by Maity et al. (2015), the north-eastern, northern and Himalayan regions of India are likely to be severely affected by stronger warming in the future. They also found that heavy rainfall in summer seasons with little change in the winter season is expected to be received by the Himalayan belt. Snow cover is considered as an active and multivariate part of the climate system (Changchun et al. 2008). Increase in temperature will also accelerate snow and glacial melt in the Subansiri Basin which in turn will increase water availability. Also, as the Subansiri Basin is highly influenced by the southwest monsoon rainfall during May to October, the climate change that results in variation in intensity of the monsoon will affect both high and low flows in the basin.

Accurate estimation of runoff is important for flood forecasting, reservoir operation, watershed management, water supply and design of hydrologic and hydraulic structures. For estimation of direct runoff from a watershed produced by a given precipitation, various models are available. The empirical models establish the relationship between rainfall and runoff based on the hydro-meteorological data and contain no physical transformation function to relate input to output (Sarkar & Kumar 2012). Artificial neural network (ANN) is one of the most simple but robust models for mapping the non-linear relation between rainfall and runoff even though it cannot represent the physical process of the catchment (Hsu et al. 1995). Thus, sometimes an ANN-based model does not give better results as the underlying physical processes have been ignored in the model. Again, rainfall alone cannot explain the runoff variance efficiently which can be attributed to antecedent moisture content (AMC), rainfall intensity, and physical characteristics of the catchment such as geology, soil, slope and land use/land cover (LULC) conditions (Costa et al. 2003). Thus, the main constraint of the empirical models is the non-consideration of physical processes such as sub-surface flow, surface runoff, and infiltration in the catchment. The physically based models, though data intensive, are able to represent the spatial variability of land surface characteristics such as topographic elevation, slope, aspect, vegetation, soil and climatic parameters such as precipitation, temperature and evapotranspiration (Niehoff et al. 2002; Akbari & Singh 2012).

Considering the advantages and disadvantages of various models, this paper presented a Soil Conservation Service (SCS) (ANN-SCS) curve number (CN) based hybrid model for estimation of runoff of the River Subansiri. A simple ANN model is also developed to evaluate the performance of the hybrid model. The ANN model uses the non-linear autoregressive model with exogenous inputs (NARX) network for obtaining the runoff. LULC and climate change can significantly influence the interception, evapotranspiration, infiltration, soil moisture, and water balance of a watershed system (Talib & Randhir 2017). Hence, along with rainfall, it is also of much importance to analyze the flow scenario of a river under different LULC characteristics. Thus, in the hybrid model, four different LULC scenarios have been created to study its impact on the future flow scenario of the river. The data derived from the Global Climate Model (GCM) have been used to generate the future rainfall over the basin. The rainfall thus predicted had been used to develop rainfall-runoff models for predicting the future runoff of the river under different scenarios. Figure 1 shows the Subansiri River basin along with the map of India and the Brahmaputra River basin.

Figure 1

Study area showing (a) India with Brahmaputra basin, (b) Brahmaputra and Subansiri basins and (c) Subansiri basin.

Figure 1

Study area showing (a) India with Brahmaputra basin, (b) Brahmaputra and Subansiri basins and (c) Subansiri basin.

MATERIALS AND METHODS

Data used

Rainfall data

The Asian Precipitation Highly Resolved Observational Data Integration towards Evaluation of Water Resources (APHRODITE) project develops high resolution daily gridded precipitation datasets for Asia (Mitra et al. 2003; Rajeevan et al. 2006). The basic algorithm adopted is based on the model developed by Xie et al. (2007). APHRODITE's long-term continental-scale product is useful for depiction of the areal distribution and variability around Himalayas, Southeast Asia and mountain regions of the Middle East (Yatagai et al. 2012).

APHRODITE's daily gridded rainfall data of spatial resolution 0.25° × 0.25° from 1960 to 2007 has been used in this study. In order to carry out spatial analysis, 24 APHRODITE grids were selected over the Subansiri basin (Figure 2).

Figure 2

The 24 different APHRODITE grids chosen for the study.

Figure 2

The 24 different APHRODITE grids chosen for the study.

GCM parameters

Monthly climate parameters of HadCM3 GCM of the A2 scenario available at a spatial resolution of 2.5° × 3.75° (latitude by longitude) have been used as the predictors. The A2 scenario represents a very heterogeneous world with continuously increasing global population. The climatic parameters taken in this study are, downward short-wave flux (dswf), near surface relative humidity (hurs), geopotential height at 200 hPa (zg200), geopotential height at 500 hPa (zg500), maximum temperature at near surface, minimum temperature at near surface, air temperature at 500 hPa (ta500), air pressure at sea level (mslp), soil moisture, precipitation rate, wind speed height above ground and relative humidity at 850 hPa (hur850). The predictors were selected based on correlation coefficient values determined by the Pearson correlation method at all 24 APHRODITE grids (Barman & Bhattacharjya 2015).

The GCM data were downloaded from the Intergovernmental Panel on Climate Change (IPCC) in two different assessments: Fourth Assessment Report (AR4) and Third Assessment Report (AR3). The time period of the Fourth Assessment is from 2001 to 2100 and the third assessment is 1890–2099. Apart from being one of the major models used in the IPCC Third (2001) and Fourth (2007) Assessments, HadCM3 also contributes to the Fifth Assessment (2013). The major advantage of the model, at the time it was developed, was its good simulation of the current climate without using flux adjustments and it still ranks highly compared with other models in this respect (Reichler & Kim 2008). Another advantage of HadCM3 is its capability to capture a time-dependent fingerprint of historical climate change in response to natural and anthropogenic forcings (Stott et al. 2000). This has made the model a particularly useful tool in studies concerning the detection and attribution of past climate changes. Moreover, HadCM3 model has the highest ‘skill scores’ for both precipitation and temperature of all the models used for the AR4 (Cai et al. 2009).

There is only one HadCM3 GCM point that falls inside the study area. Hence, three other HadCM3 points have also been considered that fall around the Subansiri basin. The geographic coordinates of these four GCM points are: 90°E, 27.5°N; 90°E, 30°N; 93.75°E, 27.5°N; and 93.75°E, 30°N (Figure 3).

Figure 3

HadCM3 GCM points that fall in and around the basin.

Figure 3

HadCM3 GCM points that fall in and around the basin.

Discharge data

Daily discharge data for the Subansiri river from 1990 to 2014 at Khabulighat gauge station were used in this study. The data was collected from the Water Resources Department, Government of Assam.

Land use/land cover data

LULC maps of the Subansiri basin have been prepared using the supervised classification method and maximum likelihood classifier algorithm using MODIS image MOD09A1.5 (MODIS/Terra Surface Reflectance 8-Day L3 Global 500 m SIN Grid) of 500 m resolution for 2002 and 2012 for the month of October. Different classes considered in preparing the LULC maps are: dense vegetation, light vegetation, bare soil, surface water bodies and snow cover area. An example LULC map of the Subansiri basin is shown in Figure 4.

Figure 4

Land use/land cover map of the Subansiri river basin, October 2008. SWB, surface water bodies; SCA, snow cover area; LV, light vegetation; DV, dense vegetation; BS, bare soil.

Figure 4

Land use/land cover map of the Subansiri river basin, October 2008. SWB, surface water bodies; SCA, snow cover area; LV, light vegetation; DV, dense vegetation; BS, bare soil.

Soil map

The USDA National Resource Conservation Service (NRCS) soil map has been used in this study. This global soil regions map is based on a reclassification of the FAO-UNESCO soil map of the world combined with a soil climate map. The map has geographic projection and its scale is 1:5,000,000. The soil map has been processed in Erdas Imagine and ArcGIS software for georeferencing and digitization. Five different soil types have been considered for the basin. These are:

  • Alfisols: high in silt size and larger particles

  • Entisols: loamy and clayey

  • Vertisols: high clay content

  • Inceptisols: very fine sand, loamy fine sand

  • Histosols: highly porous

Accordingly, the soil types are divided into the four hydrologic soil groups (HSG) as follows:

  • Inceptisols and Histosols: Group A

  • Alfisols: Group B

  • Vertisols: Group C

  • Entisols: Group D

Downscaling of rainfall using statistical downscaling techniques and its prediction

Barman & Bhattacharjya (2015) have carried out a study for predicting the future rainfall of the Subansiri basin. They considered rainfall at 24 locations over the Subansiri catchment. They used multiple linear and non-linear regression-based statistical downscaling techniques and also adopted the ANN model. These techniques were compared on the basis of mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R2), mean absolute percentage error (MAPE) and root mean square percentage error (RMSPE). The study showed that with minimum values of all these errors, the multiple non-linear regression-based downscaling technique outperformed the multiple linear regression and the ANN model. Table 1 shows the average error values for the downscaling techniques. Several studies compared the ANN model with other downscaling techniques and revealed that few other models performed better than the ANN model. For example, Gaume & Gosset (2003) compared feed-forward ANN with the linear model and conceptual model. They concluded that their conceptual model outclassed the linear and ANN models. Goyal & Ojha (2010) applied several linear regression-based downscaling models such as direct, forward, backward, and stepwise regression for downscaling mean monthly precipitation in the Pichola watershed, India. They concluded that the direct regression-based downscaling model yielded better performance among all the techniques for that region.

Table 1

Average values of MAE, RMSE, R2, MAPE and RMSPE

Downscaling methods Errors
 
MAE RMSE R2 MAPE RMSPE 
Multiple linear regression 41.88 58.33 0.79 33.81% 48.22% 
Multiple non-linear regression 37.71 55.04 0.83 29.49% 44.21% 
Artificial neural network 40.99 56.99 0.81 32.55% 46.23% 
Downscaling methods Errors
 
MAE RMSE R2 MAPE RMSPE 
Multiple linear regression 41.88 58.33 0.79 33.81% 48.22% 
Multiple non-linear regression 37.71 55.04 0.83 29.49% 44.21% 
Artificial neural network 40.99 56.99 0.81 32.55% 46.23% 

Owing to its better performance, the non-linear regression-based statistical downscaling technique has been used to predict rainfall in the Subansiri basin up to 2099 for four different time frames: namely, 2020–39, 2040–59, 2060–79 and 2080–99 at all 24 APHRODITE grids. The Mann-Kendall (M-K) (Mann 1945; Kendall 1975) test, which is a non-parametric test to identify trends in time series data, was performed in XLSTAT for each month and each time frame and the resultant Sen's slope values were used to understand the future increase and decrease of rainfall.

Runoff simulation of Subansiri

The runoff of Subansiri has been simulated using two different models up to 2099: the ANN based rainfall-runoff model; and the ANN-SCS based hybrid model. In order to incorporate the spatial variability, the entire catchment has been divided into 11 different sub-catchments. Figure 5 shows the 11 sub-catchments of the Subansiri basin along with the APHRODITE grids that fall in each sub-catchment.

Figure 5

The Subansiri basin with 11 sub-catchments.

Figure 5

The Subansiri basin with 11 sub-catchments.

Development of the ANN based rainfall-runoff model

The ANN model has emerged as the most popular method for prediction because of its ability to recognize time series patterns and non-linear characteristics (Jones 2004). In this study, the ANN model was run using the time series analysis. The method selected is the non-linear autoregressive with external (exogenous) input (NARX). NARX is a recurrent dynamic network, with feedback connections enclosing several layers of the network (Khamis & Abdullah 2014). Chang et al. (2014) developed three ANN models to make forecasts on the evolution of water level in a floodwater storage pond (FSP). The results demonstrated that the NARX has higher applicability than the back-propagation neural network (BPNN) and the Elman neural network. Fereidoon & Koch (2018) applied the NARX neural network to predict continuous rainfall series across the Karkheh river basin, Iran and found that the NARX model has a high potential for real-time rainfall prediction. Nanda et al. (2016) developed a wavelet based NARX (WNARX) model to address the issues of non-availability of real-time rainfall and discharge data and satellite-based highly biased rainfall data products for real time flood forecasting in upper Mahanadi River basin. The satellite-based rainfall products and observed rainfall products were tested for their use in linear autoregressive moving average with exogenous inputs (ARMAX), ANN, wavelet-based ANN (WANN), NARX and WNARX models. The results showed that the model performances vary in the order WNARX > NARX > WANN > ANN > ARMAX.

The NARX model is based on the linear ARX model and it predicts time series y(t) from past values of that time series and past values of another time series x(t), called the external or exogenous time series (Billings 2013). Mathematically, it can be defined as,  
formula
(1)
For the present study, y(t) is the runoff at the final outlet, x(t) is the average rainfall at each sub-catchment and d represents the tapped delay lines that store previous values of the x(t) and y(t) sequences. Monthly time series data have been used in the present study. The model was constructed using neural network toolbox in MATLAB 8.3 (R2014a). Testing, training and validation were performed for the period 1990 to 2014, for which observed discharge data are available. In the network modeling, out of the total data, 70% were selected for training and 15% each for validation and testing. NARX is a two-layer feed forward network with sigmoid transfer function in the hidden layer and linear transfer function in the output layer. The number of neurons in the hidden layer was determined by the trial and error method. Based on the coefficient of determination (R) values, the numbers of neurons in the hidden layers selected were 15 and number of delays selected was 1 as the optimum network. The network was created and trained in open loop form. Open loop allows the user to supply the network with correct past outputs as it is trained to produce the correct current outputs. Figure 6 shows the NARX neural structure for open loop. After training, the network was converted to closed loop form for multi-step prediction up to 2099. Closed loop converts neural network open-loop feedback to closed loop. Figure 7 represents the NARX neural structure for closed loop. The preparets function available in MATLAB is used to prepare the input and target time series data for network simulation or training. This function simplifies the normally complex and error-prone task of reformatting input and target time series. It automatically shifts input and target time series as many steps as are needed to fill the initial input and layer delay states. In this study, the monthly runoff has been predicted for four different time frames: namely, 2020–39, 2040–59, 2060–79 and 2080–99 taking 1990–2014 as the base period. Also, the future trends of runoff for different months and different time periods have been analyzed on the basis of Sen's slope values.
Figure 6

NARX neural structure for open loop.

Figure 6

NARX neural structure for open loop.

Figure 7

NARX neural network (closed loop).

Figure 7

NARX neural network (closed loop).

Development of an ANN-SCS-based hybrid model

Along with rainfall, the basin characteristics of a catchment such as LULC and soil properties, among others, also influence runoff. In this study, an attempt has been made to develop a hybrid model for runoff simulation. This hybrid model is based on ANN model and SCS-CN model. The empirical equation involved in the SCS approach requires rainfall and a watershed coefficient called the curve number (CN) as inputs. CN is an index that represents the combination of HSG and LULC classes. The model involves relationships between LULC, HSG and CN. Khosrashahi & Saghafian (2004) integrated the LULC and HSG maps of Damavand watershed using Integrated Land and Water Information System (ILWIS) GIS software and curve numbers were calculated for the entire basin and sub-basins. Nayak et al. (2012) used ILWIS GIS platform for preparation of LULC maps and their change detection study between 2001 and 2007 for Uri river watershed in Lower Narmada basin in Central India. They used the SCS-CN method and found that surface runoff volume increased in 2007 compared with 2001 for similar rainfall events.

In the present study, the SCS-CN model in ILWIS GIS platform has been applied to determine runoff for all 11 sub-catchments of Subansiri basin from 1990 to 2099. These runoffs were then fed as input to the ANN model to simulate the runoff of Subansiri at the final outlet. The CN maps have been created in the ILWIS software with LULC and HSG maps as inputs. The average CN value obtained from the histogram of CNMAP was considered as the CN-II for the watershed. CN-I and CN-III were then calculated using standard equations as given in SCS-CN model theory (USDA 1985). Accordingly, the potential maximum retention (S1, S2 and S3) values were also calculated. By applying the AMC conditions, the curve numbers were determined for each sub-catchment. CN-I was obtained for January, February, March, November and December, CN-II was obtained for the months of April, September and October while CN-III was obtained for May, June, July and August.

Creation of different land use/land cover scenarios

The effect of change in LULC on the flow scenario of Subansiri has been studied by considering different LULC scenarios. A LULC change detection study was performed by supervised classification method using maximum likelihood classifier algorithm in Erdas Imagine software during the period 2002 to 2012. Quantitative areal data of the overall LULC changes as well as gains and losses in each category between 2002 and 2012 were then compiled. From the analysis, considerable decreases in dense vegetation (24.66%) and snow cover area (1.3%) were observed during this period. On the other hand, increases in light vegetation (13.09%) and surface water bodies (3.46%) for the same time period were also detected. Figure 8 is the graphical representation of the changes for the entire decade. Applying this changing trend in future, the following four different scenarios have been developed for the different time periods to study the impact of change in LULC on future runoff of the Subansiri river:

  • Scenario 1: Bare land, dense vegetation and snow cover area have been reduced and were covered by light vegetation and surface water bodies during the period 2020–2039.

  • Scenario 2: Under this scenario, further reductions have been made in bare soil, dense vegetation and snow cover area for the period 2040–2059. These were then covered by light vegetation and surface water bodies.

  • Scenario 3: Under this scenario, reduction has been made on area covered by light vegetation resulting in the formation of bare soil for the period 2060–2079.

  • Scenario 4: For the period 2080–2099, in addition to the conditions made in the previous sub-scenarios, effect of melting of snow has been taken into account. Increase in surface water bodies due to melting of snow is considered under this scenario.

Figure 8

LULC change in Subansiri river basin from 2002 to 2012.

Figure 8

LULC change in Subansiri river basin from 2002 to 2012.

RESULTS AND DISCUSSION

Rainfall analyses over the Subansiri basin

The future average rainfall over the Subansiri basin have been analyzed for four different time frames. The Sen's slope values are represented as a heat map as shown in Figure 9. For most of the months, the results revealed negative (decreasing) trends for the different time periods with a maximum in July during 2060–79 followed by June during the periods 2080–99 and 2060–79 respectively. The maximum positive (increasing) trend was observed during 2060–79 for April followed by August during 2020–39 and February during 2060–79, respectively.

Figure 9

Heat map showing Sen's slope of rainfall trends for different months and for different time frames.

Figure 9

Heat map showing Sen's slope of rainfall trends for different months and for different time frames.

During 2020–39, except for June, July and September when rainfall showed a negative trend, for all other months the rainfall will have increasing trends. During the period 2040–59, rainfall shows increasing trends for January, March, April and September. Rainfall shows decreasing trends for the rest of the months. During 2060–79, positive rainfall trends were observed for February, March, April and December. Again, positive trends for the months of January, February, March, April, September and October were observed during 2080–99 while for the remaining six months, the rainfall showed negative trends. It is seen from the analysis that, in future, the basin will experience less rainfall during the monsoon and more rainfall during the non-monsoon periods. These changing patterns in rainfall over the basin can be attributed to climate change.

Runoff analysis for ANN based rainfall-runoff model

The ANN based rainfall-runoff model has been analyzed for four different future time frames. Figure 10 gives the regression plot for the study. The coefficient of correlation (R) values obtained are 0.94, 0.81 and 0.73, respectively, for the training, validation and testing phase, with an overall R value of 0.89. R values less than 0.35 represent weak correlations, 0.36 to 0.67 represent moderate correlations, and 0.68 to 1.0 represent strong correlations (Weber & Lamb 1970; Mason et al. 1983). Thus, the overall R value of 0.89 obtained for the ANN based rainfall-runoff model shows a strong correlation between observed and the simulated data. The average runoff predicted for each month for different time periods is presented as a bar diagram in Figure 11. It is observed that, except for August and September, for the remaining months, runoff for the time frame 2020–39 decreases compared with the base period 1990–2014. Again, compared with 2020–39, the runoff first increases for 2040–59 and then decreases from 2060 to 2099 for January, February, March, April, September, October, November and December. For May, June and August, runoff increases from 2020 to 2099. There is a possibility of occurrence of 5000–6000 cumec of runoff during 2060–79 and 2080–99 for June, July and August. The line diagrams for the same can be seen in Figures S1 and S2 in the supplementary information.

Figure 10

Regression plot for the ANN based rainfall-runoff model.

Figure 10

Regression plot for the ANN based rainfall-runoff model.

Figure 11

Bar diagram representing runoff of Subansiri for different time frames as obtained from the ANN based rainfall-runoff model.

Figure 11

Bar diagram representing runoff of Subansiri for different time frames as obtained from the ANN based rainfall-runoff model.

The above results were for the average runoff value of each month for the entire individual periods. However, while studying the trend of future runoff for each month by taking all the values for the entire individual periods on the basis of Sen's slope values, it is seen that in the months for which the rainfall shows decreasing trends the runoff also shows negative trends and vice versa (Figure 12). For example, Sen's slope value was at a minimum in July during 2060–79 in case of rainfall trend analysis and Sen's slope value for runoff was also found to be at the minimum for July during the same period. Similarly, in both the heat maps, for April during 2060–79, the positive trend was at its maximum. Thus, from the simple ANN model, it has been found that the higher the rainfall the greater the runoff and vice versa. Thus, it becomes necessary to study the impact of changes in climate variables that may occur in the near future on runoff. For example, a simultaneous increase in precipitation and evaporation may produce no change in the runoff even though both these parameters change (Stonevicius et al. 2016). Again, a combination of the increase in temperature and variation in precipitation will result in a significant change of the runoff (Mousavi et al. 2018).

Figure 12

Heat map showing Sen's slope of runoff trends for different months and for different time frames obtained for the ANN-based rainfall runoff model.

Figure 12

Heat map showing Sen's slope of runoff trends for different months and for different time frames obtained for the ANN-based rainfall runoff model.

Runoff analysis for the ANN-SCS based hybrid model

Considering four different LULC scenarios, the runoff of Subansiri has been determined using an ANN-SCS based hybrid model. The CN values for different scenarios are shown in Table 2.

Table 2

CN values for each sub-catchment for each of the scenarios

  SC1 SC2 SC3 SC4 SC5 SC6 SC7 SC8 SC9 SC10 SC11 
CNI 
 S1(2020–39) 57.1 56.9 50.1 42.8 51.4 50.1 75.8 75.1 47.5 60.9 44.7 
 S2(2040–59) 51.7 52.8 49.4 47.9 51.2 49.4 74.6 74.6 48.3 57.1 44.9 
 S3(2060–79) 62.1 60.4 55.1 57.8 58.2 55.1 74.6 74.6 54.6 57.6 51.3 
 S4(2080–99) 58.4 57.7 52.3 55.2 55.2 52.3 73.1 73.1 51.3 57.6 51.3 
CNII 
 S1(2020–39) 76 75.9 70.5 64.1 71.6 70.5 88.2 87.8 68.3 78.8 65.8 
 S2(2040–59) 71.8 72.8 69.8 68.7 71.4 69.8 87.5 87.5 69 76 66 
 S3(2060–79) 79.6 78.4 74.5 76.6 76.8 74.5 87.5 87.5 74.1 76.4 71.5 
 S4(2080–99) 77 76.5 72.3 74.6 74.6 72.3 86.6 86.6 71.5 76.4 71.5 
CNIII 
 S1(2020–39) 87.9 87.9 84.6 80.4 85.3 84.6 94.5 94.3 83.2 89.5 81.6 
 S2(2040–59) 85.4 85.9 84.2 83.5 85.2 84.2 94.2 94.2 83.6 87.9 81.7 
 S3(2060–79) 89.9 89.3 87.0 88.2 88.2 87.0 94.1 94.1 86.8 88.1 85.2 
 S4(2080–99) 88.5 88.2 85.7 87.1 87.1 85.7 93.6 93.6 85.2 88.1 85.2 
  SC1 SC2 SC3 SC4 SC5 SC6 SC7 SC8 SC9 SC10 SC11 
CNI 
 S1(2020–39) 57.1 56.9 50.1 42.8 51.4 50.1 75.8 75.1 47.5 60.9 44.7 
 S2(2040–59) 51.7 52.8 49.4 47.9 51.2 49.4 74.6 74.6 48.3 57.1 44.9 
 S3(2060–79) 62.1 60.4 55.1 57.8 58.2 55.1 74.6 74.6 54.6 57.6 51.3 
 S4(2080–99) 58.4 57.7 52.3 55.2 55.2 52.3 73.1 73.1 51.3 57.6 51.3 
CNII 
 S1(2020–39) 76 75.9 70.5 64.1 71.6 70.5 88.2 87.8 68.3 78.8 65.8 
 S2(2040–59) 71.8 72.8 69.8 68.7 71.4 69.8 87.5 87.5 69 76 66 
 S3(2060–79) 79.6 78.4 74.5 76.6 76.8 74.5 87.5 87.5 74.1 76.4 71.5 
 S4(2080–99) 77 76.5 72.3 74.6 74.6 72.3 86.6 86.6 71.5 76.4 71.5 
CNIII 
 S1(2020–39) 87.9 87.9 84.6 80.4 85.3 84.6 94.5 94.3 83.2 89.5 81.6 
 S2(2040–59) 85.4 85.9 84.2 83.5 85.2 84.2 94.2 94.2 83.6 87.9 81.7 
 S3(2060–79) 89.9 89.3 87.0 88.2 88.2 87.0 94.1 94.1 86.8 88.1 85.2 
 S4(2080–99) 88.5 88.2 85.7 87.1 87.1 85.7 93.6 93.6 85.2 88.1 85.2 

SC, sub-catchment; S, scenario.

The runoff at each sub-catchment was determined by the SCS-CN method. Then the runoff at the final outlet was calculated using ANN time series analysis and the NARX method. The regression plots are shown in Figure 13. The coefficient of correlation (R) values obtained for training, validation and testing are 0.96, 0.89 and 0.92, respectively, with an overall R value of 0.93. This indicates that the model is competent enough to be used for simulation of runoff of Subansiri. Considering 1990 to 2014 runoff as the base, the future monthly runoff has been predicted and analyzed for the scenarios created for the four different future time periods. The bar diagram representing these future runoff changes is shown in Figure 14. The results show that except for the months of April and August, for all months, the runoff first decreases during 2020–39 compared with the base period 1990–2014 (Figure 14). For January and February, runoff first decreases during 2040–59 compared with 2020–39 and then increases again for the next two periods. Runoff increases from 2020 to 2099 for March, April, May, June, July, August and November. For September and October, runoff first increases compared with 2020–39 and then decreases again from 2060 to 2099. For December, the runoff decreases for the period 2040–59 compared with 2020–39 and then again gradually increases for the remaining periods. For scenarios 1, 2, 3 and 4, the maximum runoff was observed in the months of June, July, July and August, respectively. The minimum runoff for the first two scenarios was observed in March while for the next two sub-scenarios, it was observed in November. The line diagrams for the same can be seen in Figures S3 and S4 in the supplementary information.

Figure 13

Regression plots for ANN-SCS based hybrid model.

Figure 13

Regression plots for ANN-SCS based hybrid model.

Figure 14

Future runoff obtained from four different scenarios.

Figure 14

Future runoff obtained from four different scenarios.

A heat map has been prepared to understand the change in trends of runoff under these four different LULC scenarios (Figure 15). It is seen that except for the months of June, July, August and September during 2020–39, October during 2040–59, November during 2060–79, and February, April and October during 2080–99, all other months show increasing trends. The maximum positive trend is seen for the month of July during 2080–99 followed by May during 2040–59 and July during 2060–79, respectively.

Figure 15

Heat map showing Sen's slope of runoff trends for different months and for different time frames obtained for the ANN-SCS based hybrid model.

Figure 15

Heat map showing Sen's slope of runoff trends for different months and for different time frames obtained for the ANN-SCS based hybrid model.

The increase and decrease of runoff at various time periods is primarily because of the change in the LULC pattern. In the different scenarios, the dense vegetation has been reduced and covered by light vegetation accompanied by urbanization. Urbanization is generally the most forceful of all land use changes potentially affecting the hydrology of an area (Leopold 1968). As the land development occurs, infiltration capacities within the watershed decrease due to the increase in impervious area, which results in a greater volume of runoff (Gilroy & McCuen 2012). Urbanization leads to increased air temperature at the surface as well as changes in the spatial pattern and intensities of precipitation (Giannaros et al. 2013). Thus, runoff in a basin is sensitive to the changes in both climate and LULC. Precipitation alone is not sufficient to explain the change in discharge trend in a watershed. The speed of climate change has the potential to significantly alter the river flow (Musau et al. 2015) and LULC changes are accelerating due to population growth and human-induced economic developments, thereby affecting the interception, infiltration and evaporation processes of the hydrologic cycle (Zeng et al. 2013). As a result, the generation and distribution of streamflow is deeply affected by the changes in precipitation, temperature and evapotranspiration in the hydrological process (Jha et al. 2007). Also, due to global warming, the snow cover area at the upper Subansiri basin is gradually decreasing and this diminishing role of snow will have an impact on the runoff downstream.

Comparison of ANN based rainfall-runoff model and ANN-SCS based hybrid model

The training, validation and testing of both the ANN models used to simulate the runoff of the Subansiri have been performed by the time series analysis in MATLAB software. As given in Table 3, the R values for training, validation and testing for the ANN-SCS-based hybrid model are 0.96, 0.89 and 0.92, which are better than the R values obtained for the simple ANN-based rainfall runoff model (0.93, 0.81 and 0.72 for training, validation and testing, respectively). The overall R value for the hybrid model is 0.93 which is better than the value of 0.89 obtained for the simple ANN model. This implies that the ANN-SCS-based hybrid model is a better performing model and could efficiently be used to simulate river runoff.

Table 3

R values for the two different runoff simulation models

Model R values
 
Training Validation Testing All R 
ANN based rainfall-runoff model 0.93 0.81 0.72 0.89 
ANN-SCS-based hybrid model 0.96 0.89 0.92 0.93 
Model R values
 
Training Validation Testing All R 
ANN based rainfall-runoff model 0.93 0.81 0.72 0.89 
ANN-SCS-based hybrid model 0.96 0.89 0.92 0.93 

CONCLUSIONS

In this study, the runoff of the Subansiri River at Khabulighat outlet has been simulated using two different models. The entire Subansiri basin has been divided into 11 different sub-catchments and the average rainfall at each sub-catchment has been determined. These average rainfall values were then fed as the input to the ANN-based rainfall-runoff model for predicting the runoff of the basin at the outlet. An ANN-SCS-based hybrid model has been developed to study the impact of change in LULC on future runoff. The runoffs at 11 different sub-catchments have been determined using the SCS-CN method. The performance of the models was assessed on the basis of R values. The ANN-SCS based hybrid model gave better R values than the simple ANN based rainfall-runoff model. There was better simulation efficiency of the ANN-SCS based hybrid model. The incorporation of the LULC information to the ANN model has significantly improved the performance of the ANN-SCS model. The ANN-SCS model is then used to evaluate the effect of LULC on the basin discharge. Four different scenarios based on LULC were created. This evaluation shows that the future runoff of the basin will be greatly impacted by the change in LULC of the basin.

REFERENCES

REFERENCES
Akbari
S.
&
Singh
R.
2012
Hydrological modelling of catchments using MIKE SHE
.
IEEE-International Conference on Advances in Engineering, Science and Management
(ICAESM-2012)
,
Nagapattinam, Tamil Nadu
,
335
340
.
Arnell
N. W.
1999
Climate change and global water resources
.
Glob. Env. Change
9
,
31
49
.
Barman
S.
&
Bhattacharjya
R. K.
2015
Comparison of linear regression, non-linear regression and artificial neural network model for downscaling of rainfall at Subansiri River Basin, Assam, India
.
European Water
51
,
53
64
.
Billings
S. A.
2013
Nonlinear System Identification: NARMAX Methods in Time, Frequency and Spatio-Temporal Domains
.
Wiley
. ISBN: 978-1-119-94359-4.
Cai
X.
,
Wang
D.
,
Zhu
T.
&
Ringler
C.
2009
Assessing the regional variability of GCM simulations
.
Geophysical Research Letters
36
,
L02706
.
doi: 10.1029/2008GL036443
.
Chang
F. J.
,
Chen
P. A.
,
Lu
Y. R.
,
Huang
E.
&
Chang
K. Y.
2014
Real-time multi-step-ahead water level forecasting by recurrent neural networks for urban flood control
.
J. Hydrol.
517
,
836
846
.
Changchun
X.
,
Yaning
C.
,
Weihong
L.
,
Yapeng
C.
&
Hongtao
G.
2008
Potential impact of climate change on snow cover area in the Tarim River Basin
.
Environ. Geol.
53
,
1465
1474
.
doi: 10.1007/s00254-007-0755-1
.
Costa
M. H.
,
Botta
A.
&
Cardille
J. A.
2003
Effects of large-scale changes in land cover on the discharge of the Tocantins River, Southeastern Amazonia
.
Journal of Hydrology
283
,
206
217
.
Gaume
E.
&
Gosset
R.
2003
Over-parameterization, a major obstacle to the use of artificial neural networks in hydrology
.
Hydrology and Earth System Sciences
7
(
5
),
693
706
.
Giannaros
T. M.
,
Melas
D.
,
Daglis
I. A.
,
Keramitsoglou
I.
&
Kourtidis
K.
2013
Numerical study of the urban heat island over Athens (Greece) with the WRF model
.
Atmospheric Environment
73
,
103
111
.
Gilroy
K. L.
&
McCuen
R. H.
2012
A non-stationary flood frequency analysis method to adjust for future climate change and urbanization
.
Journal of Hydrology
414–415
,
40
48
.
Gogoi
C.
,
Goswami
D. C.
&
Phukan
S.
2013
Flood risk zone mapping of the Subansiri sub-basin in Assam, India
.
International Journal of Geomatics and Geosciences
4
(
1
),
75
88
.
Goyal
M. R.
&
Ojha
C. S. P.
2010
Evaluation of various linear regression methods for downscaling of mean monthly precipitation in arid Pichola watershed
.
Nat. Resour.
1
,
11
18
.
Hsu
K. L.
,
Gupta
H. V.
&
Sorooshian
S.
1995
Artificial neural network modeling of the rainfall–runoff process
.
Water Resour. Res.
31
,
2517
2530
.
IPCC
2001
Chapter 8: Model evaluation
. In:
Climate Change 2001: The Scientific Basis. Contribution of Working Group I to the Third Assessment Report of the Intergovernmental Panel on Climate Change
(
Houghton
J. T.
,
Ding
Y.
,
Griggs
D. J.
,
Noguer
M.
,
van der Linden
P. J.
,
Dai
X.
,
Maskel
K.
&
Johnson
C. A.
, eds).
Cambridge University Press
,
Cambridge
,
UK
and New York
, pp.
471
523
.
IPCC
2007
Chapter 8: Climate models and their evaluation
. In:
Contribution of Working Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change
(
Solomon
S.
,
Qin
D.
,
Manning
M.
,
Chen
Z.
,
Marquis
M.
,
Averyt
K. B.
,
Tignor
M.
&
Miller
H. L.
, eds).
Cambridge University Press
,
Cambridge
,
UK
and New York
, pp.
590
662
.
IPCC
2013
Chapter 9: Evaluation of climate models
. In:
Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change
(
Stocker
T. F.
,
Qin
D.
,
Plattner
G.-K.
,
Tignor
M.
,
Allen
S. K.
,
Boschung
J.
,
Nauels
A.
,
Xia
Y.
,
Bex
V.
&
Midgley
P. M.
, eds).
Cambridge University Press
,
Cambridge
,
UK
and New York
, pp.
741
866
.
IRG Systems South Asia Pvt Ltd
2014
Cumulative Impact and Carrying Capacity Study of Subansiri sub Basin Including Downstream Impacts
.
Submitted to Central Water Commission, Ministry of Water Resources
,
New Delhi
,
India
,
Final report, Vol.-II
.
Jha
M.
,
Arnold
J. G.
,
Gassman
P. W.
,
Giorgi
F.
&
Gu
R. R.
2007
Climate change sensitivity assessment on upper Mississippi River basin streamflows using SWAT
.
J. Am. Water Resour. Ass.
42
,
997
1015
.
Jones
A. J.
2004
New tools in non-linear modelling and prediction
.
Computational Management Science
1
,
109
149
.
Kendall
M. G.
1975
Rank Correlation Measures
.
Charles Griffin
,
London
.
Khamis
A.
&
Abdullah
S. N. S. B.
2014
Forecasting wheat price using back propagation and NARX neural network
.
Int. J Eng. Sci. (IJES)
3
(
11
),
19
26
.
Khosrashahi
M.
&
Saghafian
B.
2002
Dynamics involved in the identification and resolution of flood prone areas in watershed
. In:
Proceedings of the Sixth International Conference on River Engineering, Shahid Chamran University
, pp.
1375
1383
.
Leopold
L. B.
1968
Hydrology for Urban Land Planning
.
USGS Circular 554
.
US Geological Survey
,
Reston, Virginia
.
Maity
R.
,
Aggarwal
A.
&
Chanda
K.
2015
Do CMIP5 models hint at a warmer and wetter India in the 21st century?
Journal of Water and Climate Change
7
(
2
),
280
295
.
Mann
H. B.
1945
Non-parametric tests against trend
.
Econometrica
13
,
245
259
.
Mason
R. O.
,
Lind
D. A.
&
Marchal
W. G.
1983
Statistics: An Introduction
.
Harcourt Brace Jovanovich, Inc
,
New York
, pp.
368
383
.
Mitra
A. K.
,
Das Gupta
M.
,
Singh
S. V.
&
Krishnamurti
T. N.
2003
Daily rainfall for the Indian monsoon region from merged satellite and rain gauge values: large scale analysis from real time data
.
J. Hydrometeorology
4
,
769
781
.
Musau
J.
,
Sang
J.
,
Gathenya
J.
&
Luedeling
E.
2015
Hydrological responses to climate change in Mt. Elgon watersheds
.
J. Hydrol. Reg. Stud.
3
,
233
246
.
Nayak
T.
,
Verma
M. K.
&
Hema
B. S.
2012
SCS curve number method in Narmada Basin
.
International Journal of Geomatics and Geosciences
3
(
1
),
219
228
.
Rajeevan
M.
,
Bhate
J.
,
Kale
J. D.
&
Lal
B.
2006
High resolution daily gridded rainfall data for the Indian region: analysis of break and active monsoon spells
.
Current Science
91
,
296
306
.
Reichler
T.
&
Kim
J.
2008
How well do couple models simulate today's climate?
Bull. Amer. Meteor. Soc.
89
,
303
311
.
Sarkar
A.
&
Kumar
R.
2012
Artificial neural networks for event based rainfall–runoff modeling
.
Journal of Water Resource and Protection
04
,
891
897
.
Stonevicius
E.
,
Rimkus
E.
,
Stara
A.
,
Kazys
J.
&
Valiukskevicius
G.
2016
Climate change impact on the Nemunas River basin hydrology in the 21st century
.
Boreal Environment Research
22
,
49
65
.
Stott
P. A.
,
Tett
S. F. B.
,
Jones
G. S.
,
Allen
M. R.
,
Mitchell
J. F. B.
&
Jenkins
G. J.
2000
External control of 20th century temperature by natural and anthropogenic forcings
.
Science
290
,
2133
2137
.
Talib
A.
&
Randhir
T. O.
2017
Climate change and land use impacts on hydrologic processes of watershed systems
.
Journal of Water and Climate Change
8
(
3
),
363
374
.
USDA
1985
National Engineering Handbook. Section 4: Hydrology, Chapter 4. Soil Conservation Service
,
USDA
,
Washington, D.C.
Weber
J. C.
&
Lamb
D. R.
1970
Statistics and Research in Physical Education
,
Vol. 222
.
CV Mosby Co
,
St. Louis
, pp.
59
64
.
Xie
P.
,
Yatagai
A.
,
Chen
M.
,
Hayasaka
T.
,
Fukushima
Y.
,
Liu
C.
&
Yang
S.
2007
A gauge-based analysis of daily precipitation over East Asia
.
Journal of Hydrometeorology
8
,
607
627
.
Yatagai
A.
,
Kemiguchi
K.
,
Arakawa
O.
,
Hamada
A.
,
Yasutomi
N.
&
Kitoh
A.
2012
APHRODITE: Constructing long-term daily gridded precipitation dataset for Asia based on a dense network of rain gauges
.
Bulletin of the American Meteorological Society
published online:
1401
1415
.
doi:10.1175/BAMS-D-11-00122.1
.

Supplementary data