Abstract

Climate changes, as well as land cover changes, affect the flow regimes in streams. Understanding the contributions of climate variables and land cover changes on low flows will help planners and decision makers to improve water resources management. An approach which uses data driven artificial neural networks (ANNs) is proposed in this study. Land cover, rainfall, snow and temperature were used as inputs to the ANN model. In this approach, an index called relative strength effect was used to assess the contribution of each input used in ANN. The proposed approach was experimented in three contrasting watersheds in northwest Indiana, USA. The study indicates that the changes in low flow regime for a less urbanized watershed were explained by land cover changes up to 30% while the remaining 70% variations were explained by meteorological inputs. In the watershed with a more developed area, the low flow variations were influenced up to 80% by meteorological inputs.

INTRODUCTION

Low flow regime is an essential component for water resources management, such as low flow augmentation, domestic water supply, irrigation, supporting living organisms, nutrient balance management and control sea water intrusion (Greco & Larsen 2014). Apart from management strategies, changing watershed land cover, as well as climate variables, contribute to the low flow variations. Understanding the influence of each component will help the managers in improved decision making.

The World Meteorological Organization (WMO 1974) defines low flow in a stream as the average flow during dry weather periods. Low flow of a river is a seasonal characteristic and is usually defined by the upper bound called mean runoff in the considered time scale (example: Mean Annual Runoff (MAR), Mean Monthly Flow (MMF) or Mean Daily Flow (MDF)) (Smakhtin 2001). Very low flow is also expressed using probability such as 7Q10 (7 day 10 year low flow defined by USGS) flow for design criteria. When we use flow duration data, low flow days are defined as days with stream flow lower than flow magnitude with the select probability of exceedance such as 80 or 85% (Lins & Slack 1999).

Many recent studies indicated the increasing as well as decreasing trends in low flow regimes in different parts of the world. Abdul-Aziz & Burn (2006) used the Mann–Kendall (MK test) non-parametric test with a trend free pre-whitening approach to examine the Mackenzie River Basin in northern Canada. They indicated a strong increasing trend in annual minimum flows in the winter months. Schilling & Libra (2003) also indicated the same fact and attributed it to the presence of the large agricultural land use pattern in Iowa, USA, where crop land covers 70% of the land area. Whitehead et al. (2015) presented the impacts of climate change on the low flow regime and the sea water intrusion in Brahmaputra and Meghna river systems in India. Baltas et al. (2010) examined the impacts of the change in rainfall regimes on surface and groundwater storages in an experimental watershed. This study concluded that the change in the rainfall conditions has a direct influence on both surface and groundwater storages. Douglas et al. (2000) presented a detailed analysis on floods as well as low flows and proposed a revised MK test, called the regional average Kendall test, to examine the trends. Lins & Slack (1999) indicated the increasing trends in low flow regime in United States rivers in general. In that study, they considered 395 stream gage stations and examined the quantiles of discharge with 0–100th percentile. They reported significant trends in low flow (Q0) and median flow (Q50) categories.

To evaluate changes in low and high flow regimes, efforts were previously made to relate the influence of watershed land cover changes and climate changes. Stohlgren et al. (1998) analyzed the variations in the stream flow pattern due to land cover modifications in a watershed. Allan (2004) presented the difficulties in finding empirical associations to study the influence of land use changes on stream flow and stream ecosystems. Thanapakpawin et al. (2007) explored the effects of land cover change in Chiang Mai, Thailand. They used a distributed hydrology soil-vegetation model to simulate the hydrologic response of the considered system and assessed the influence of upland shifting cultivation. Nie et al. (2011) developed a soil and water assessment tool (SWAT) for the upper San Pedro watershed and studied the impact of land cover changes on different hydrological components to help the stakeholders in decision making. These studies showed the complex nature of the interrelationship between flow regime and land use changes.

In this work, attempts were made to explain the variations in the low flow regime using an artificial neural network (ANN) to help the planners to use site-specific data driven model in their decision making. Artificial neural network models were very widely used by many researchers in river flow prediction, water resources management and drought analysis (Jain et al. 2004). A detailed review of ANN models is available in Maier & Dandy (2000), Maier et al. (2010) and ASCE (2000). The data driven models are universal functional approximators and can capture interrelationships by training the network using training algorithms (Maier et al. 2010). Due to its capacity to capture nonlinear interrelationships, in this study a simple feed forward three layered neural network was used to model the complex process. After successful training, the ANN models can also be used to study the influence of each input in the output variations. Relative Strength Effect (RSE), introduced by Kim et al. (2001), was used as in index in the past for this purpose (Song & Chandramouli 2013). ANNs were used to study the influence of different land cover types and the climate variables over the low flow regime in the northwest Indiana.

DATA SOURCE

Daily flow, mean monthly values of temperature, precipitation, snow, and land cover data were used in the model building process. Two contrasting watersheds called Deep River and Little Calumet East Arm from northwest Indiana, USA were considered for this study (Figure 1). The Deep River system and Hart Ditch has more developed (45 and 49% respectively) area than the Little Calumet East Arm watershed (17% developed) (Table 1). Deep River and Hart Ditch drain to the Great Lake Michigan through Burns Ditch. Hart Ditch drains to Lake Michigan through Cul-De Sag canal. Data from LaPorte, Valparaiso, and Lowell meteorological stations located in northwest Indiana were used due to their proximity to the watersheds considered. Daily flow observations (source: Unites States Geological Survey (USGS)), rainfall, snowfall, temperature and land cover (source: National Oceanographic Atmospheric Agency (NOAA) national climatic data center) were downloaded from the public domain. Processed land use data is available for every 5 years from 1996 onwards (source: NOAA Digital Coast data center). After basic geoprocessing using ARC GIS software, the land cover data for 1996, 2001, 2006 and 2011 for the two watersheds were extracted and consolidated.

Table 1

Watershed details

Details Deep River watershed Little Calumet East Arm Hart Ditch 
Area (km2465.9 191.5 183.1 
Major watershed Lake Michigan Lake Michigan Lake Michigan 
Developed area (2011) (km2209.7 34.4 88.9 
Agriculture (2011) (km2154.0 70.5 55.0 
Forest (2011) (km257.2 55.9 25.6 
Wetland (2011) (km245.0 30.7 13.6 
% change in developed area (compared with 2006 land cover data) 45% 18% 49% 
Flow characteristics 
Mean (cfs) 60 80 76 
Standard dev. (cfs) 37 28 40 
Record length (year) 63 65 68 
Q with probability of exceedance 95%) (cfs) 17 31 
Q with probability of exceedance 90%) (cfs) 22 34 10 
Q with probability of exceedance 80%) (cfs) 30 38 15 
Details Deep River watershed Little Calumet East Arm Hart Ditch 
Area (km2465.9 191.5 183.1 
Major watershed Lake Michigan Lake Michigan Lake Michigan 
Developed area (2011) (km2209.7 34.4 88.9 
Agriculture (2011) (km2154.0 70.5 55.0 
Forest (2011) (km257.2 55.9 25.6 
Wetland (2011) (km245.0 30.7 13.6 
% change in developed area (compared with 2006 land cover data) 45% 18% 49% 
Flow characteristics 
Mean (cfs) 60 80 76 
Standard dev. (cfs) 37 28 40 
Record length (year) 63 65 68 
Q with probability of exceedance 95%) (cfs) 17 31 
Q with probability of exceedance 90%) (cfs) 22 34 10 
Q with probability of exceedance 80%) (cfs) 30 38 15 
Figure 1

Watersheds considered.

Figure 1

Watersheds considered.

INITIAL EXAMINATION OF FLOW TRENDS

Before setting up the ANN model, efforts were initially taken to study the shifts in flow regimes in the recent three decades using flow duration curves. Flow duration curves (Fennessey & Vogel 1990) of the recent three decades shifted upwards in the low flow regime when compared to the previous three decades for both watersheds (Figures 2 and 3). Since this method disturbs the chronological sequence, it can be used only for initial examination. Subsequently, by considering the entire time span of data, trends were examined for their statistical significance using the MK test.

Figure 2

Little Calumet East Arm: flow duration curves.

Figure 2

Little Calumet East Arm: flow duration curves.

Figure 3

Deep River: flow duration curves.

Figure 3

Deep River: flow duration curves.

The statistical validity of the shifts in low flow trends were examined using the MK test. The non-parametric MK test uses chronologic data (Zhang et al. 2001). The null hypothesis H for the MK test is defined as ‘there is no trend in the data series’ (Zhang et al. 2001; Kumar et al. 2009). The MK test statistics (S) are defined in Equation (1):  
formula
(1)
where n = total number of points in the dataset.
The variance of S depends on ‘n’ and is defined as:  
formula
(2)
Z is calculated as given in Equation (2) as recommended in Hirsch & Slack (1984) using the normal distribution. S indicates the nature of the trend. When S is zero, no trend exists. Positive and negative values of S indicate increasing and declining trends, respectively. For examining the trends in low flow, a new monthly series was constructed from daily flow data. From 70 years of daily data, monthly mean flows were estimated initially. The number of days in a month with flow magnitude less than the mean monthly flow (Fi) is estimated (Equation (3)):  
formula
(3)

Using this monthly frequency time series, trends in low flow were examined using the MK test for each month. For both the watersheds, the MK test statistics (S) were positive, indicating increasing trends. However, they were not statistically significant at the 95% confidence level (Table 2). Similarly, the rainfall and temperature, as well as snowfall, did not show a statistically significant trend. After this check, for both watersheds, observed data were used in the ANN model.

Table 2

Monthly low flow data results

Station name No of data Trend observed Confidence level (%) 
Little Calumet East Arm 478 845 Increasing (but not statistically significant) 70 
Deep River 478 545 Increasing (but not statistically significant) 65 
Station name No of data Trend observed Confidence level (%) 
Little Calumet East Arm 478 845 Increasing (but not statistically significant) 70 
Deep River 478 545 Increasing (but not statistically significant) 65 

ANN MODEL

An ANN model was defined for the monthly time scale. Fi for a month (Equation (3)) was defined as the output of the ANN model. Monthly rainfall (RF), snowfall (S), temperature (T), lag 1, lag 2 and lag 3 values of RF, S and T were given as inputs. Along with that, land cover details were also given as inputs. From the extracted land cover data from 1996 to 2011 (Figure 4), monthly land cover values were interpolated linearly. The coastal land cover given by NOAA is in the C-CAP (Coastwatch Change Analysis Project) classification system with 26 land cover types (Klemas et al. 1993). Extracted land cover data were lumped into three inputs, namely Developed (D), Agricultural (AG) and Forest (FR) groups. Each group had a 10–15% change in the 15 year span. These changes may have contributed to the low flow regime shift. As the change in forest cover area was minimal, it was not used in the model. D, AG, D Lag1 (DL1), D Lag 2 (DL2), AG Lag 1 (AGL1) and AG Lag 2 (AGL2) were given as inputs along with meteorological variables. In total, 18 inputs were considered in the stage 1 ANN model (Figure 5). Each input was normalized as indicated in Melesse et al. (2011) using maximum and minimum values of that input series (Equation (5)):

 
formula
(4)
Figure 4

Land cover details for LCE and Deep River watersheds.

Figure 4

Land cover details for LCE and Deep River watersheds.

Figure 5

ANN model.

Figure 5

ANN model.

In the ANN model construction, to study the land use and climate inputs and their significance, discharges were not used as inputs because the target output was derived from daily flow series.

Using the Marquardt Levenberg algorithm (Kisi & Uncuoglu 2005), the ANN model was trained. Out of 178 monthly datasets, 80% were used for training and 20% were used for testing and validation. By monitoring the mean square error and mean relative error (ASCE 2000) of the testing dataset, optimal termination was performed. RSE, introduced by Kim et al. (2001), was used to relate the importance of each input used in the ANN model. This approach was successfully used in the past (Brion et al. 2005; Ramalingam & Lingireddy 2014). After completing the training, RSE was calculated for each input.

RSE BACKGROUND

RSE of an input is defined as the partial derivative of the output. By defining hidden and output layer neurons with logistic activation functions, RSE can be found as:  
formula
(5)
where = max{||, = 1 to I}; = number of datasets considered for training; I = number of inputs; = value of weight between the ith input neuron and jth hidden neuron; = value of weight between the jth hidden neuron and kth output neuron; ; ; = threshold value at the kth output neuron; and = threshold value at the jth hidden neuron.
Song & Chandramouli (2013) presented the details of finding RSE in a feed forward ANN model. Influential input will have an RSE value close to 1 (for a positively correlated input) or −1 (for a negatively correlated input) after an optimal training. However, for less influential inputs, RSE will be close to zero (0.1 to −0.1). Hence, RSE values vary between 1 and −1. RSE values were brought to a comparable percentage scale as SRSE using the following scheme:  
formula
(6)

RESULTS AND DISCUSSION

Using SRSE values, the influence of each input over the output was interpreted in this section. In general, the influence of different variables over the output can be interpreted from basic processes. The rainfall has a positive correlation with low flow as well as high flow regimes (Todd & Mays 2005). According to the output defined in this case study (number of days with discharge less than mean monthly discharge), rainfall should have a negative correlation with the output. On the other hand, more water loss could occur in the watershed system due to temperature increase (through evaporation and evapo-transpiration). So, the temperature could have a positive correlation with the output. Output is expected to get less influence from snow during the month because snow melt can influence low flow mostly with time lags. Inferences derived from ANN results follow this pattern mostly, but it showed deviations for a few inputs.

Deep River watershed

From the Stage 1 model with 18 inputs (Figure 6), Stage 2 model inputs were selected. Inputs with less contribution to output were identified using SRSE and removed from the Stage 1 model. Five inputs which had SRSE very close to zero (TL2, T, DL1, AG, AGL2) were eliminated for Stage 2. Stage 2 ANN model was constructed using 13 inputs and another two inputs (RF3, DL2) were eliminated using the same criteria for the final model.

Figure 6

DR watershed Stage 1 ANN model results-SRSE values.

Figure 6

DR watershed Stage 1 ANN model results-SRSE values.

The final Stage 3 model with 11 inputs (hereafter called DR ANN model) was used to understand the contribution of meteorological variables and land use variables on the output (Figure 7).

Figure 7

DR watershed Stage 3 ANN model results-SRSE values.

Figure 7

DR watershed Stage 3 ANN model results-SRSE values.

RF, RFL1 and RFL2 have a strong influence on the output in both watersheds. Negative values of SRSE for those inputs indicate that when rainfall increases, output decreases. RF alone explains 35% of the output variations. SL2 and SL3 also influence the output in a similar manner like rainfall. An increase in the TL1 makes the output increase and it explains the more than 10% variations of the output.

Snow during the month, as well as SL1, had a positive RSE. This indicates that the snow melt contribution has a time lag to the low flow. TL3 input has negative SRSE. This is due to the combined influence of TL3 with other inputs such as rainfall and snow.

D input has a negative influence to the output. Historic data from these two watersheds show a negative correlation between D and the output. ANN captured this information. Ag input with lag 1 influences the output negatively. For the Deep River watershed, meteorological variables explain 80% of the low flow variations and the remaining 20% is explained by the land cover variables.

LCE watershed

The LCE watershed has a less developed area than the Deep River watershed. Similar to the DR watershed, the Stage 1 ANN model for LCE watershed was also initiated with 18 inputs. However, the Stage 2 and final models were decided using SRSE based elimination. Based on the SRSE values of the Stage 1 model (Figure 8), six variables, namely RFL3, TL3, SL3, SL2, SL1 and TEMP, were eliminated to construct the Stage 2 model.

Figure 8

LCE watershed Stage 1 ANN model results-SRSE values.

Figure 8

LCE watershed Stage 1 ANN model results-SRSE values.

In the final LCE model, 10 inputs were used (Figure 9) after eliminating two more inputs in the second stage. The final model for LCE and DR watersheds were similar in rainfall input influences. For the LCE watershed, rainfall alone explains 47.5% of the output variations. In the SRSE based elimination process, snow, snow with lag1, lag 2 and lag 3 were eliminated due to their minimal contributions. Temperature lag 1 and lag 2 had positive SRSE explaining 22% of the output variations. Similar to Deep River, developed land cover input has a negative influence over the output (19%). However, AgL1 and AgL2 has a positive influence on the output. Overall, meteorological and land cover variables contribute 69.5 and 30.5% respectively to the output variations.

Figure 9

LCE watershed Stage 3 ANN model results-SRSE values.

Figure 9

LCE watershed Stage 3 ANN model results-SRSE values.

Hart Ditch watershed

The third watershed called Hart Ditch was also considered to verify the results derived from the other two watersheds. Hart Ditch is similar to Deep River because it has a larger developed area (49%). Like the other two watersheds, a Stage 1 model was developed using 18 inputs. The final model for Hart Ditch had 15 inputs and the SRSE for each input is shown in Table 3. For the Hart Ditch watershed, land cover variables explain the variations of output up to 28%.

Table 3

SRSE values of Hart Ditch ANN model inputs

Inputs SRSE (%) 
RF2 −3.77 
RF1 −9.97 
TL2 −4.03 
TL1 11.13 
SL3 4.10 
SL1 5.20 
Temp 5.41 
Rain −20.80 
Snow 7.34 
−4.52 
AG 6.07 
DL1 −3.95 
AGL1 5.55 
DL2 −3.26 
AGL2 4.91 
Inputs SRSE (%) 
RF2 −3.77 
RF1 −9.97 
TL2 −4.03 
TL1 11.13 
SL3 4.10 
SL1 5.20 
Temp 5.41 
Rain −20.80 
Snow 7.34 
−4.52 
AG 6.07 
DL1 −3.95 
AGL1 5.55 
DL2 −3.26 
AGL2 4.91 

These case studies very clearly show the effect of developed area as a critical factor, as indicated by Ahn & Merwade (2017). The watershed low flow regime is less explained by land cover if the watershed has a more developed area. All the watersheds had different influential inputs as well as having varying contributions from inputs. Seventy to 80% of the variations in low flow were controlled by inputs with meteorological variables. Among the three watersheds, the Deep River watershed is twice as large as the other two watersheds. In the Deep River watershed, which had larger developed land cover, the low flow variations were controlled significantly by meteorological variables. For the LCE watershed, land cover change influences the output more. The Hart Ditch watershed also had a large developed area and it had 28% contributions from land cover. So, the input contributions to the output variations are very site specific.

Examining very low flow regime

In the earlier sections, outputs to define low flow regime were decided using the mean flow (50%) criteria (Equation (3)). In this section, efforts were made to redefine the low flow regime as ‘very low flow regime’ using a lower critical value for each month. The Deep River watershed was considered for this analysis. For each month, with Weibull's plotting position formula, probability of exceedance values were established using daily flows. Using the probability of exceedance as 80%, monthly critical values were decided (Table 4). If a daily flow in a month is less than this value, it is considered as a very low flow hit. Number of days in a month with flow less than the critical value (hits in a month) were found for each month. These monthly frequencies were defined as the target for the ANN model. Using 18 inputs the Stage 1 model was developed. Using SRSE based elimination, the Stage 3 model for the very low flow regime had 11 inputs (Figure 10). Very low flow regime is highly influenced by rainfall and lagged rainfall inputs. RF, RF1, RF2 and RF3 together contribute 48% of the very low flow regime and land cover changes contribute 26%.

Table 4

Monthly critical values for Deep River watershed for low flow and very flow regimes

Month Critical low flow magnitude defined using average flow (50% probability of exceedance) Critical very low flow magnitude defined using 80% probability of exceedance 
January 136 25 
February 146 30 
March 219 70 
April 212 60 
May 169 43 
June 132 26 
July 75 16 
August 66 14 
September 67 12 
October 63 14 
November 87 21 
December 121 25 
Month Critical low flow magnitude defined using average flow (50% probability of exceedance) Critical very low flow magnitude defined using 80% probability of exceedance 
January 136 25 
February 146 30 
March 219 70 
April 212 60 
May 169 43 
June 132 26 
July 75 16 
August 66 14 
September 67 12 
October 63 14 
November 87 21 
December 121 25 
Figure 10

Very low flow regime–Deep River final ANN model.

Figure 10

Very low flow regime–Deep River final ANN model.

Limitations

The proposed ANN modeling is very effective for site specific analysis. These models are data intensive. For hydrologically similar watersheds, in regional specific applications, useful details can be derived by this method. In the two case studies, interpolated monthly land cover data were used as inputs. As the land cover changes were occurring near the urban boundaries steadily, this approximation was adopted. Also, the case study watersheds were considered from the Lake Michigan watershed. Due to the proximity to huge freshwater resource, inputs such as the water demand, water use for domestic and agricultural sectors were not considered. However, these inputs can also be included in the ANN model together with other inputs and their influence over the output can be estimated easily using SRSE. Meteorological variables are random, but land cover data are controllable by planners. By varying the land cover area, optimal values of land cover to be maintained can be estimated using the trained ANN. This will help the planners to control the rapid changes in land use by implementing best management practices to augment low flow regimes as well as reservoir operations.

CONCLUSIONS

Identifying the influences of land cover changes and meteorological variables on the low flow regime in a watershed scale always remained as a challenge for planners. The ANN based data driven model proposed in this study provides a promising way to study these influences. Three case studies used in this study indicate the influence of meteorological variables as predominant (70–80%). Twenty to 30% of low flow variations were explained by land cover variables, which is considerable. For a watershed with a more developed area, the meteorological variables were highly influential. On the other hand, for a watershed with a less developed area, the land cover data explains 30% of low flow variations. When very low flow regimes were modeled using ANN, land cover influences contributed 28% and the rainfall related inputs were highly influential in the model.

ACKNOWLEDGEMENTS

This research work was supported by the Indiana Department of Environmental Management, Department of Natural Resources Coastal Grants (CZ 324).

REFERENCES

REFERENCES
Abdul-Aziz
,
O. I.
&
Burn
,
D. H.
2006
Trends and variability in the hydrological regime of the Mackenzie River basin
.
J. Hydrol.
319
(
1–4
),
282
294
.
Baltas
,
E. A.
,
Dervos
,
N. A.
&
Mimikou
,
M. A.
2010
Impact of changing rainfall conditions on surface and ground water resources in an experimental watershed in Greece
.
Global NEST J.
12
(
2
),
119
125
.
Brion
,
G.
,
Viswanathan
,
C.
,
Neelakantan
,
T. R.
,
Lingireddy
,
S.
,
Girones
,
R.
,
Lees
,
D.
,
Allard
,
A.
&
Vantarakis
,
A.
2005
Artificial neural network prediction of viruses in shellfish
.
Appl. Environ. Microbiol.
71
(
9
),
5244
5253
.
Douglas
,
E. M.
,
Vogel
,
R. M.
&
Kroll
,
C. N.
2000
Trends in floods and low flows in the United States: impact of spatial correlation
.
J. Hydrol.
240
,
90
105
.
Fennessey
,
N.
&
Vogel
,
R. M.
1990
Regional flow-duration curves for ungauged sites in Massachusetts
.
ASCE J. Water Resour. Plan. Manage.
116
(
4
),
530
549
.
Hirsch
,
R. M.
&
Slack
,
J. R.
1984
A nonparametric trend test for seasonal data with serial dependence
.
Water Resour. Res.
20
(
6
),
727
732
.
Jain
,
A.
,
Sudheer
,
K. P.
&
Srinivasulu
,
S.
2004
Identification of physical processes inherent in artificial neural network rainfall runoff models
.
Hydrol. Process.
18
(
3
),
571
581
.
Kim
,
C. Y.
,
Bae
,
G. J.
,
Hong
,
S. W.
,
Park
,
C. H.
,
Moon
,
H. K.
&
Shin
,
H. S.
2001
Neural network based prediction of ground surface settlements due to tunneling
.
Comput. Geotech.
28
(
6–7
),
517
547
.
Kisi
,
O.
&
Uncuoglu
,
E.
2005
Comparison of three backpropagation training algorithms for two case studies
.
Indian J. Engng Mater. Sci.
12
,
443
450
.
Klemas
,
V. V.
,
Dobson
,
J. E.
,
Furguson
,
R. L.
&
Haddad
,
K. D.
1993
A coastal land cover classification system for the NOAA Coastwatch Change Analysis Project
.
J. Coastal Res.
9
,
862
872
.
Kumar
,
S.
,
Merwade
,
V.
,
Kam
,
J.
&
Thurner
,
K.
2009
Streamflow trends in Indiana effects of long term persistence, precipitation and subsurface drains
.
J. Hydrol.
374
(
1–2
),
171
183
.
Lins
,
H. F.
&
Slack
,
J. R.
1999
Streamflow trends in the United States
.
Geophys. Res. Lett.
26
,
227
230
.
Melesse
,
A. M.
,
Ahmad
,
S.
,
McClain
,
M. E.
,
Wang
,
X.
&
Lim
,
Y. H.
2011
Suspended sediment load prediction of river systems: an artificial neural network approach
.
Agric. Water Manage.
98
,
855
866
.
Nie
,
W.
,
Yuan
,
Y.
,
Kepner
,
W.
,
Nash
,
M. S.
,
Jackson
,
M.
&
Erickson
,
C.
2011
Assessing impacts of landuse and landcover changes on hydrology for the upper San Pedro watershed
.
J. Hydrol.
407
,
105
114
.
Ramalingam
,
D.
&
Lingireddy
,
S.
2014
Neural network-derived heuristic framework for sizing surge vessels
.
J. Water Resour. Plan. Manage.
140
(
5
),
678
692
.
Schilling
,
K. E.
&
Libra
,
R. D.
2003
Increased baseflow in Iowa over the second half of the 20th century
.
J. Am. Water Resour. Assoc.
39
(
4
),
851
860
.
Smakhtin
,
V. U.
2001
Low flow hydrology: a review
.
J. Hydrol.
240
,
147
186
.
Stohlgren
,
T. J.
,
Chase
,
T. N.
,
Pielke
,
R. A.
,
Kittel
,
G. F. T.
&
Baron
,
J. S.
1998
Evidence that local land use practices influence regional climate, vegetation and stream flow patterns in adjacent natural areas
.
Glob. Chang. Biol.
4
(
5
),
495
504
.
Thanapakpawin
,
P.
,
Richey
,
J.
,
Thomas
,
D.
,
Rodda
,
S.
,
Campbell
,
B.
&
Logsdon
,
M.
2007
Effect of landuse change on the hydrologic regime of the Mae Chiang river basin, NW Thailand
.
J. Hydrol.
334
(
1–2
),
215
230
.
Todd
,
D. K.
&
Mays
,
L. W.
2005
Groundwater Hydrology
.
John Wiley Inc.
,
NJ
,
USA
.
Whitehead
,
P. G.
,
Barbour
,
E.
,
Futter
,
M. N.
,
Sarkar
,
S.
,
Rodda
,
H.
,
Caesar
,
J.
,
Butterfield
,
D.
,
Jin
,
L.
,
Sinha
,
R.
,
Nicholls
,
R.
&
Salehin
,
M.
2015
Impacts of climate change and socio-economic scenarios on flow and water quality of the Ganges, Brahmaputra and Meghna (GBM) river systems: low flow and flood statistics
.
Environ Sci Process Impacts
17
(
6
),
1057
1069
.
doi:10.1039/c4em00619d
.
WMO (World Meteorological Organization)
1974
International Glossary of Hydrology
.
WMO
,
Geneva
.
Zhang
,
X.
,
Harvey
,
K. D.
,
Hogg
,
W. D.
&
Yuzyk
,
T. R.
2001
Trends in Canadian streamflow
.
Water Resour. Res.
37
(
4
),
987
998
.