Modeling Hydrologic and Water Quality Responses to Changing Climate and Land Use/Cover in the Wolf Bay Watershed, South Alabama by Ruoyu Wang A thesis submitted to the Graduate Faculty of Auburn University in partial fulfillment of the requirements for the Degree of Master of Science Auburn, Alabama December 13, 2010 Keywords: Climate change, LULC change, SWAT, Flow, Sediment, TN, TP Copyright 2010 by Ruoyu Wang Approved by Latif Kalin, Chair, Assistant Professor of School of Forestry and Wildlife Sciences Hanqin Tian, Alumni Professor of School of Forestry and Wildlife Sciences Xing Fang, Associate Professor of Department of Civil Engineering ii Abstract Land use/cover (LULC) and climate change are two main factors directly affecting regional hydrology and water quality. In this study, the future potential impacts of LULC and climate change on the hydrologic regimes and water quality in Wolf Bay watershed, South Alabama were explored independently and mutually by using the Soil and Water Assessment Tool (SWAT). Due to lack of measured data, SWAT was calibrated in a nearby watershed, and the calibrated model parameters were transferred to the Wolf Bay watershed. It was shown that using data from nearby watersheds improves the model performance under limited data conditions in the study watershed. The choice of the parameter set, whether it is the default model parameters or those from a donor watershed, has a marginal effect on modeling the impacts of different LULC scenarios. SWAT with the transferred parameters was then employed to investigate the potential impacts of LULC and climate change on the hydrology and water quality of the Wolf Bay watershed. While four Global Circulation Models (GCMs) under three Green House gas emission scenarios were used to reflect variability in future climate conditions, three future LULC maps generated mainly based on different population growth rate assumptions were used to represent the uncertainty in future LULC conditions. In general, the Wolf Bay watershed is expected to experience increasing precipitation in the future, especially in fall, and temperature is expected to be higher, especially in summer and fall months. Further, the watershed is expected to undergo dramatic urbanization, with percentage of urban areas nearly doubling in future. iii Results showed that both climate change and LULC change would cause a redistribution of streamflow. Higher flows were projected to increase, while small flows are expected to decrease. No clear trend of extreme large flow was detected when only climate change was considered. Under combined change scenarios, a more noticeable uneven distribution of streamflow was observed. Monthly average streamflow was projected to increase in spring, fall, and winter, especially during the fall, while no clear trend was observed in summer. LULC change did not significantly affect monthly streamflow, but changed the partitioning of streamflow to baseflow and surface runoff. Surface runoff was predicted to increase every month, while for baseflow an evident decreasing trend was detected. When climate was combined with LULC effect, a more dramatic increasing trend in monthly average streamflows was detected. Furthermore, a visible increasing trend in surface runoff and more dramatic decreasing trend in baseflow were detected. Monthly distribution of sediment and nutrients are affected by both flow and management practices. Projected variations of TSS, TN, and TP loadings follow the same pattern as flow. No evident difference in annual average N:P ratio was predicted when only climate change was considered. LULC change increased TSS loadings but decreased TN loadings for all months. TP loadings were projected to decrease in summer, but increase in other months. N:P ratio was projected to decrease significantly. Results of this study indicate that if future loadings are expected/predicted to increase/decrease under either climate or LULC change scenario, then their combined impact is to intensify that trend. On the other hand, if their effects are in opposite directions, that is while one predicts an increase and the other predicts a decrease, then their mutual effect has an offsetting impact. The combined LULC and climate change effect was in general synergistic, i.e. the total effect was greater than the sum of the individual effects. iv Acknowledgments I owe my deepest gratitude to my major professor Dr. Latif Kalin. Without his encouragement this work would not have been possible. In addition, I would like to sincerely thank my graduate committee members, Dr. Hanqin Tian and Dr. Xing Fang for their steadfast support and dedication to the completion of this thesis. Thanks are due to Shufen Pan?s group for providing land use/cover map of 2005. I also want to thank Dr. Mingliang Liu for his help in GCM data and Dr. Sabahattin Isik for helping me in programming. I would also like acknowledge my friend and officemate Rewati Niraula for his help during this research, and other colleagues, Amir Sharifi, Andrew Morrison, Harsh Singh and Nishan Bhattarai. Thanks are due to Auburn University?s Water Resources Center for funding my project. Finally, to my girlfriend and family who have supported me tirelessly throughout this project, I express my most heartfelt appreciation. Without your love and laughter I would have not been capable of completing this work. v Table of Contents Abstract ......................................................................................................................................... ii Acknowledgments........................................................................................................................ iv List of Tables ............................................................................................................................... vi List of Figures ............................................................................................................................. vii Chapter I: Modeling Effects of Land Use/Cover Changes under Limited Data ........................... 1 Abstract ............................................................................................................................. 1 Introduction ....................................................................................................................... 2 Methodology ..................................................................................................................... 6 Results and discussion .................................................................................................... 11 Summary and conclusions .............................................................................................. 20 References ....................................................................................................................... 22 Chapter II: Responses of Hydrological Processes and Water quality to LULC and Climate Change in Wolf Bay Watershed, Southern Alabama ........................... 39 Abstract ........................................................................................................................... 39 Introduction ..................................................................................................................... 40 Methodology ................................................................................................................... 45 Results and discussion .................................................................................................... 55 Summary and conclusions .............................................................................................. 72 References ....................................................................................................................... 76 Chapter III: Summary and Conclusions .................................................................................... 103 vi List of Tables CHAPTER I Table 1 Physical similarities between Wolf Bay and Magnolia River watersheds .................... 27 Table 2 Calibrated SWAT parameters (flow part) and their default values ............................... 27 Table 3 Model performance (flow part) at Magnolia River watershed ...................................... 28 Table 4 Model performance (water quality part) at Magnolia River watershed ......................... 28 Table 5 Calibrated SWAT parameters (water quality part) and their default values .................. 29 Table 6 Model performance at daily time scale in Wolf Bay watershed .................................... 29 Table 7 Kolmogorov?Smirnov tests for daily and monthly simulated streamflows generated with different parameter sets. .................................................... 29 Table 8 Land use/cover (LULC) and impervious area (IA) change in Wolf Bay watershed ...................................................................................... 30 Table 9 Annual statistics of streamflow under different LULC conditions................................ 30 CHAPTER II Table 1 Land use/cover in Wolf Bay watershed for baseline (2005) and 3 future scenarios (2030) ......................................................... 83 vii List of Figures CHAPTER I Fig. 1 Geographical location of Wolf Bay and Magnolia River watersheds .............................. 31 Fig. 2 Simulated streamflow compared with observed data from 1999 to 2009 in Magnolia river watershed .............................................................. 32 Fig. 3 Simulated monthly sediment and nutrient compared with observed data (LOADEST) from 2000 to 2001 in Magnolia River watershed ............................................................. 33 Fig. 4 Monthly simulation streamflow for different parameter sets in Wolf Bay watershed ............................................................................. 34 Fig. 5 Flow duration curve using different parameter sets in Wolf Bay watershed ................... 35 Fig. 6 Average monthly flow before and after land use change ................................................. 36 Fig. 7 Flow duration curve using different land use maps in Wolf Bay watershed by (a)default (b)transferred (c)calibrated parameters ..................... 37 Fig. 8 Relative change in simulated streamflows due to different LULC for different parameters in Wolf Bay watershed (1999-2009) ......................................... 38 CHAPTER II Fig. 1 Geographical location of Wolf Bay and Magnolia River watersheds .............................. 84 Fig. 2 Land use maps for Wolf Bay watershed ........................................................................... 85 Fig. 3 Seasonal mean temperature and precipitation variation from baseline (1984-2008) period according to 4 GCMs under 3 emission scenarios in wolf Bay watershed .......................................................... 86 Fig. 4 Exceedance probabilities of daily precipitation in Wolf Bay watershed .......................... 87 Fig. 5 Monthly responses of flow and respective relative changes from the baseline (only climate change effect) ................................................................ 88 viii Fig. 6 a.25-years flow duration curves(FDC) under projected future climate and current(baseline) climate, and b. Relative changes of future FDCs from the baseline (only climate change effect) ................................................................ 89 Fig. 7 Seasonal future and baseline FDCs, and seasonal relative changes of future FDCs from the baseline period (only climate change effect) ............................ 90 Fig. 8 Monthly responses of TSS, TN, TP and respective relative change from the baseline period (only climate change effect) ..................................................... 91 Fig. 9 Monthly responses of organic and mineral nutrient, and respective relative change from the baseline period (only climate change effect) ............................ 92 Fig. 10 Monthly responses of flow and respective relative change from the baseline LULC (only LULC change effect) ..................................................... 93 Fig. 11 a. 25-years flow duration curves (FDCs) under projected future LULC and current (baseline) LULC, and b. Relative changes of future FDCs from the baseline (only LULC change effect) ............................................................................................. 94 Fig. 12 Monthly responses of TSS, TN, TP and respective relative change from baseline (only LULC change effect) ............................................................................................. 95 Fig. 13 Monthly responses of organic and mineral nutrient, and respective relative changes from the baseline (only LULC change effect) ................................................................ 96 Fig. 14 Monthly responses of flow and respective relative change from the baseline (combined change effect) ................................................................... 97 Fig. 15 a.25-years flow duration curves (FDCs) under projected future situations and current(baseline) situation, and b. Relative changes of future FDCs from the baseline (combined change effect) ............................................................................ 98 Fig. 16 Seasonal future and baseline FDCs; seasonal relative changes of future FDCs from the baseline (combined change effect) ................................................................... 99 Fig. 17 Monthly responses of TSS, TN, TP and respective relative changes from the baseline (combined change effect) ................................................................. 100 Fig. 18 Monthly responses of organic and mineral nutrient, and respective relative changes from the baseline LULC (only LULC change effect) ................................................... 101 Fig. 19 LULC, climate, combined and synergic effect on average monthly percentage change of (a) streamflow, (b) TSS, (c) TP, (d) TN ....................................................... 102 1 Chapter I Modeling Effects of Land Use/Cover Changes under Limited Data ABSTRACT Watershed models are valuable tools used in the study of impacts of land use/cover (LULC) changes on hydrology. We use the Soil and Water Assessment Tool (SWAT) to study the impacts of LULC changes in a coastal Alabama watershed, where flow data did not exist at the onset of the study. We set up and calibrated the model in the neighboring Magnolia River watershed. Relevant model parameters were then transferred to the Wolf Bay watershed. Impacts of LULC changes on hydrology are studied in the Wolf Bay watershed by running the model with the default parameters, transferred model parameters (from the Magnolia River watershed), and calibrated parameters at the Wolf Bay watershed with limited data that became available later during the study. The relative changes in flow duration curves (FDCs) due to differing LULC showed a similar pattern with each parameter set: There is a clear threshold of around 1% probability of exceedance where the relative change is at its maximum. The relative change in flow due to LULC change drops drastically with increasing probability of exceedance of beyond 2% until it reaches a plateau at p D 20%. Hence, small to medium range flows are less sensitive to the parameter set. Further, the impact of LULC change on flow gradually decreases with the size of the storm for very large events (probability of exceedance <1%). 2 INTRODUCTION Quantifying the impacts of land use and land cover (LULC) changes on the hydrologic processes and water balance of river basin has been an area of interest to hydrologists in recent years. Little is known so far if there is a well defined quantitative relationship between the LULC properties and the runoff generation mechanisms. The assessment of future LULC changes with respect to their hydrological impacts is still an unsolved problem (Fohrer, 2002). Several methods were developed to study the implications of LULC changes on hydrologic processes, such as the paired catchments approach, time series analysis (statistical method) and hydrological modeling (Li et al. 2009). Among these approaches, hydrological modeling has been widely applied in many different places in the world since it requires less resource and provides more flexibility. Fohrer et al. (2001) assessed the hydrologic response to LULC changes in four meso-scale watersheds in Germany with different LULC distributions. Then the model performance for changing LULC has been tested in an artificial watershed with a single crop at a time and one underlying soil type to eliminate the complex interactions of natural watersheds. Simulation results showed that LULC changes on the annual water balance was moderate. Surface runoff was most susceptible to LULC change at both the artificial and the natural catchment. Hundecha and Bardoosy (2004) simulated the effect of LULC changes on the runoff generation of Rhine River Basin through parameter regionalization of Hydrologiska Byr?ns Vattenbalansavdelning (HBV) model. Results suggested that increased urbanization leads to an increase in the smaller peak runoffs stemming from summer storms. Increase in the larger peaks resulting from winter rainfall was negligible. A considerable reduction of both the peak runoff and the total runoff volume resulted from intensified afforestation. Savary et al. (2009) assessed the effects of 3 historical LULC change on runoff and low-flow using the Gestion Int?gr?e des Bassins versants ? l?aide d?un Syst?me Informatis? (GIBSI) model in the Chaudiere River Watershed, Canada. Simulations showed strong correlations between LULC changes and stream discharge at the outlet of the watershed, especially for summer and fall seasons. Simulated annual and seasonal low flows were also strongly correlated to agricultural and forested land. Guo et al. (2008) studied the combined effects of climate and LULC change on hydrological processes using the Soil and Water Assessment Tool (SWAT) in Poyan Lake basin, China. They found that climate effect is dominant to alter annual streamflow; while LULC change may have a moderate impact on annual streamflow. Both of them strongly influences seasonal streamflow and alter the annual hydrograph of the basin. Ma et al. (2009) also considered climate change impacts on hydrological responses in a different watershed in southwestern China by SWAT. Contrasting to the results of Guo et al. (2008), they found climate having a more profound effect on seasonal variations in streamflow with LULC change having a moderate impact. On the other hand, they observed a much stronger influence by LULC change on mean annual streamflow. Their simulation results also showed that the impact of climate change on surface water, baseflow and streamflow was offset by the impact of LULC changes. As mentioned above, LULC impacts on hydrologic responses have been thoroughly studied through modeling. However, models are mathematical simplifications of natural processes, with inevitable errors and deficits. Therefore, the reliability of hydrologic models should be evaluated by the fitness between measured flow data and model simulations. In this regard observed data is quite valuable. Hydrologists often need to adjust model variables in order to attain close to optimal parameter values by minimizing the error between model simulations and observed data. However, observed data are sometimes insufficient or not available at all, in which case one can 4 run the model without calibration by estimating parameter values from the literature or rely on regionalization approaches. The term regionalization has its roots in the process of regime classification and watershed grouping. It has later been extended in the rainfall-runoff modeling context to refer to the transfer of parameters from neighboring gauged watersheds (also called donor watersheds) to an ungauged watershed. Nowadays, the concept of regionalization applies to all methods aimed at estimating model parameter values on any ungauged watershed in a definable region of consistent hydrological response. Several methods are available in the literature for the transferring of model parameters. Regionalization based on regression is the most popular method which tries to link parameter values to climate and watershed physical characteristics, such as annual rainfall, temperature, area, slope, and land use/cover (LULC) in a gauged watershed (Yokoo, et.al, 2001; Kim and Kaluarachchi, 2008). Another commonly used approach is regionalization based on physical similarity. Generally information is transferred between neighboring watersheds, not necessarily geographically connected but rather in terms of observable watershed descriptions (Oudin et al ., 2008). Parameters are transferred from one or many donor watersheds, whose physical descriptors are similar to the ungauged one, based on one a synthetic rank that reflects the similarity of all physical descriptors between donors and target. The third kind of regionalization is based on spatial proximity. It uses the parameter values calibrated in nearby watersheds, which have sufficiently long data for calibration. The rationale of this method is that physical and climatic characteristics are relatively homogeneous within a small region, thus the neighbors should have similar hydrology. Over the past few decades, several researchers have attempted to identify the best regionalization approach appropriate for different hydrological models. For example Oudin et al. 5 (2008) applied two lumped rainfall-runoff models to daily data over a large set of 913 French catchments. Their research indicated that the spatial proximity approach provided the best solution and the regression approach was the least satisfactory in France, where a dense network of gauging station is available. Merz and Bloschl (2004) investigated the water balance dynamics of 308 catchments in Austria using the HBV model. They compared regionalization methods for estimating model parameters in ungauged catchments. The method based on multiple regressions with catchment attributes performed significantly poorer than the other two. They found spatial proximity being a better surrogate of unknown controls on runoff dynamics than catchment attributes. Reichl and Western (2009) compared Nash-Sutcliffe efficiency and monthly relative volume error of the SimHyd lumped conceptual rainfall-runoff model by averaging method, spatial proximity approach, local calibration and simple regression in 184 Australian catchments. Averaging method, which selects a number of candidate models from available gauged catchments and weighs them based on likelihood, can be considered as the improvement of regionalization by physical similarity. Their research showed that the averaging method, while inferior to local calibration, is superior to methods based on regression and spatial proximity. This paper focuses on estimating the impacts of LULC changes on hydrological responses in a coastal Alabama watershed. In particular, it investigates how limited hydrological data affects our understanding of LULC change impacts on hydrology by using the SWAT model. To address this issue, LULC maps corresponding to two different periods (1992 and 2005) are utilized. Model parameters are obtained from a nearby watershed through regionalization based on spatial proximity. Model efficiency is compared through use of time series of flow and flow duration curves (FDC) when transferred parameters and default ones are utilized. The effects of 6 parameter transferring on modeling the impacts of LULC changes on low, medium and high flows are discussed. METHODOLOGY Study Area Wolf Bay is located on the Gulf of Mexico in Baldwin County, Alabama, nestled between Pensacola Bay to the east and Mobile Bay to the west, with a watershed covering about 126 km2. It is a sub-estuary of Perdido Bay with a connection to the Intracoastal Waterway and includes various freshwater, nutrient and sediment inputs from several sub-watersheds through Wolf, Sandy, Miflin and Hammock creeks (Fig. 1). The watershed is primarily rural, but several municipalities exist including Foley, Elberta, Gulf Shores, and Orange Beach. Baldwin County?s beaches, bays and rivers promote an expanding tourism industry, which exerts substantial influences on water extraction for human uses. Baldwin County experienced a 43% increase in population from 1990 to 2000. As a result of population growth, there is an increased demand for commercial, residential, and infrastructure development, thus bringing growth management issues to the forefront for local elected officials. One of the more visible changes in the landscape of Baldwin County is the rapid transformation of agricultural and forested lands to residential development. These development pressures are threatening the natural resources which make Baldwin County a popular place to live and visit (Stallman et.al, 2005). As a result, detecting the impact of potential LULC changes is urgent and necessary, because it provides policy makers some valuable suggestion which strike a balance between development and the protection of natural resources. There is only one flow monitoring station in the watershed on the Wolf Creek operated by U.S. Geological Survey (USGS). However, at the commencement of this study no flow data was 7 available yet. USGS essentially monitors flow stage and converts them to discharge through stage-discharge curves only when they have enough flow measurements that cover range of flows. Magnolia River watershed, which is adjacent to Wolf Bay watershed to the northeast (Fig. 1), has 10 years of continuous flow and climate data. Using the regionalization based on spatial proximity, we can setup a model in Magnolia River watershed, calibrate it and transfer the model parameters to Wolf Bay watershed. Besides their spatial proximity, Wolf Bay and Magnolia River watersheds also have quite similar physical characteristics (Table 1). Although it was still partly provisional, almost 2 years of flow data (12/5/2007 to 9/30/2009) later became available from the USGS gauge on Wolf Bay Creek, which provided us an opportunity to assess the feasibility of parameter transferring from Magnolia River watershed to Wolf Bay watershed. Watershed Model SWAT is one of the most commonly used watershed models for assessing the impact of management practices and land disturbances on watershed responses. It has a solid track record of applications (Kalin and Hantush, 2006). SWAT has been widely used around the world, such as the Cottonwood River near new Ulm, Minnesota (Hanratt and Stefan, 1998); southern Alberta, Canada (Chanasyk et al., 2003); Jeker river basin, Belgium (Nasr et al., 2005) to assess various impacts of agricultural practices and land use activities on water quantity and quality. SWAT is also suitable for coastal and flat areas, which have more complicated geo-hydrologic conditions (Wu and Xu, 2006). ArcSWAT version 2.3.4 that runs on ArcGIS? was used for preparing the input data and processing the output files. SWAT is a distributed, process-based watershed model, but with significant number of empirical relationships. The physical backbone of the model facilitates the interpretation of 8 model parameters whereas the empirical simplifications keep data requirements low compared to physically based models (Heuvelmans, 2004). SWAT divides a watershed into several subwatersheds based upon drainage areas of the tributaries. Then, each subwatershed is split into multiple hydrological response units (HRUs) based on LULC and soil types. Each HRU is assumed to be spatially uniform in LULC, soil, topography, and climate. Major hydrologic process that can be simulated by SWAT include evapotranspiration (ET), surface runoff, infiltration, percolation, shallow aquifer and deep aquifer flow, and channel routing (Arnold et al., 1998). Details and the theoretical background of the SWAT is beyond the scope of this paper and can be found in Neitsch et al. (2005). In addition to streamflow, SWAT can also provide baseflow and surface runoff estimates as model outputs. Therefore, we used a baseflow filter to split the observed streamflow into baseflow and surface components to better calibrate the model. The algorithm presented by Arnold et al. (1995) is employed for this purpose. In this algorithm, a digital filter, which is borrowed from signal processing, is successively applied to streamflow. Filtering surface runoff (high frequency signals) from base flow (low frequency signals) is analogous to the filtering of high frequency signals in signal analysis and processing. The filter can be passed over the streamflow three times. At each pass, a slower component of streamflow (less baseflow as a percentage of total streamflow) is obtained. Model Performance Evaluation The statistical measures of mass balance error (MBE), coefficient of determination (R2) and Nash-Sutcliffe (1970) efficiency (ENS) are used as indicators of model performance: 9 obs obssim N i iobs N i iobs N i isim Q QQ Q QQ MBE ?= ? = ? ?? = == 1 , 1 , 1 , ])(][)([ )]()([ 1 2 , 1 2 , , 1 , 2 ?? ? == ? ?? ?? = N i simisim N i obsiobs simisim N i obsiobs OOOO OOOO R ])([ ])([ 1 1 2 , 2 1 ,, ? ? = = ? ? ?= N i obsiobs N i iobsisim NS OO OO E where Qsim,i and Qobs,i are simulated and observed flows at ith observation, respectively, N is the number of observations. Similarly, simO and obsO are average of simulated and observed flows over the simulation period. R2 describes the proportion of the total variances in the observed data that can be explained by the model and ranges from 0 to 1. ENS is a measure of how well the plot of observed versus predicted values fit the 1:1 line, and can theoretically vary from ?? to 1, with 1 denoting a perfect model with respect to data agreement. Although R2 and MBE values have been used often in the past to quantitatively compare model results with data, ENS is a better representative measure for model goodness-of-fit (ASCE 1993, Legates and McCabe 1999). Modeling LULC changes LULC changes affect various components of the hydrologic cycle, either directly or indirectly. The infiltration and ET processes are the two vital components of the hydrologic cycle directly affected by LULC changes. SWAT uses SCS curve number method to simulate infiltration process. Each soil/LULC combinations are assigned specific curve numbers, with 10 higher values representing higher surface runoff and less infiltration. Urbanization within a watershed increases the area of impervious surfaces (high Curve Number) which decreases infiltration and increases runoff. As a result, the amount of surface runoff generated from a specific rain event increases. Reduced infiltration results in less groundwater recharge which decreases baseflow contribution to streamflow, eventually causing reduction in low-flows. If change in ET is relatively small, then urbanization in essence redistributes baseflow and runoff components of the streamflow. SWAT calculates ET from potential ET (PET). One key component in PET calculation is the net radiation, which is a function of the plant albedo (reflectivity). Thus, change in LULC should change net radiation and eventually PET. In SWAT changing LULC has little or no effect on PET depending on the choice of the PET calculation method (Penman Monteith, Priestley-Taylor, and Hargreaves). In calculating the actual ET, SWAT evaporates intercepted water in the canopy first. If water intercepted in the canopy cannot fulfill the PET demand (usually the case), SWAT then calculates transpiration from plants. Transpiration is function of PET, leaf area index (LAI), and soil water content. LAI changes with land cover and plant growing seasons. Higher LAI means more transpiration. Calculation of transpiration and water uptake are described in detail in Neitsch et al. (2005). Two LULC maps representing the years 1992 and 2005 are employed to investigate the impacts of LULC changes on hydrologic responses in Wolf Bay watershed. The 1992 National Land Cover Data (NLCD) is a raster data set with a 30 m resolution. The second LULC map is produced by GIS specialists in Auburn University using circa 2005 as references data. Circa 2005 is a vector dataset attained by trend analysis focused on LULC changes of urban and built- up areas, utilities, and transportation from 2001 to 2005 based on Color Infrared imagery of 2001 11 and 2005 from Baldwin county commissions. Since these two maps had different LULC classifications, we reclassified them according to SWAT classification to make it consistent with model?s own database. RESULTS AND DISCUSSION Calibration and validation in the Magnolia River Watershed SWAT model was first set up in the Magnolia River watershed, then calibrated and validated with a split data set approach. The period from 10/01/1999 to 09/30/2004 of the daily flow data from USGS gauge #02378300 was used for calibration and the period from 10/01/2004 to 09/30/2009 was used for validation. Model validation is defined as the process of demonstrating that model is capable of making accurate predictions for periods outside a calibration period. Usually, calibration of a model is based on 3-5 years of data (Sorooshian et al. 1983; Xia et al. 2004), and validation on another period of similar length (Tu 2009, Ma et al. 2009). Table 2 shows the calibrated model parameters along with their default values. Model simulations actually started from 10/01/1989 with measured precipitation data as input. This corresponds to a warm-up period of 10 years. The idea behind using such a long warm-up period was to minimize the effect of initial unknown conditions such as antecedent moisture, and initial groundwater table height (Kalin and Hantush, 2006). Model parameters were calibrated first at monthly, then at daily time scales for flow. Fig. 2a shows the observed and simulated monthly flows during the calibration and validation periods. Monthly streamflow values match well to the observed ones. Model performance statistics are shown in Table 3. Note that only MBE is shown for baseflow as suggested by Santhi et al. (2001). It is difficult to estimate the spatial and temporal distribution of ground water table. Quantifying 12 the impact of deep aquifer system on baseflow response is also challenging (Lee et al. 2005). Therefore, it is hard to capture the temporal dynamics of baseflow simulations. Overall SWAT?s performance at monthly time scale is good during both calibration and validation periods. Daily simulations of total streamflow are not as good as monthly simulations, but the ENS of the calibration period is still acceptable. Due to the temporal scale effect discussed in the previous paragraph we only focus on total streamflow at daily time scale. According to Moriasi et al. (2007) ENS values above 0.5 with low MBE are considered satisfactory. To gain more insight we also compared FDCs of observed and simulated flows in the Magnolia River Watershed from 1999 to 2009 (Fig. 2b). Observed and simulated flows have good agreement for flows having probability of exceedance > 0.2%. For the larger flows model underestimates flow as much as 50%, which is not uncommon in modeling (e.g. Baffaut and Benson, 2009; Larose et al., 2007; Wang and Melesse, 2005). Note that SWAT is not an event-based model. Although it works reasonably well for long term simulations, it has limitations in extreme events. It cannot capture the dynamics at sub-daily scale. For example, from 31st March 2005 to 6th April 2005, there were series of several very big storms. The total amount of rainfall in this one-week period was 440 mm, which is about one fourth of the average annual precipitation. The model failed to reflect these huge events properly. The MBE of streamflow in this period was -53%. The most improper simulation happened on 1st April 2005. Observed daily average flow was 197m3/s (largest ever recorded), yet SWAT estimated only 35m3/s of flow. Such extreme events can significantly alter the performance statistics. For instance, if we ignore the event on 1st April, 2005, The ENS for monthly simulation improves from 0.65 to 0.74 (see Table 3). Other than the potential deficiencies of the model in dealing with such huge events, there are two other possible reasons for this. USGS measures 13 stage not discharge; discharge is estimated from stage-discharge relationships (i.e. regression equations) which are known to have problems outside their range. Thus, observed flow during such an extreme event, which is actually estimated from stage, could have serious errors. Spatial variation in precipitation and the rain gauges not being able to capture these accurately is another source of error. Our precipitation data source is a rain gauge located at the watershed outlet. On 1st April 2000, the USGS gauge at Magnolia River recorded a storm event where average daily flow was 6 m3/s, up from 0.6 m3/s from the day before. However, no flow is generated by SWAT because the rain gauge did not record any trace of rainfall. The most likely scenario is that it only rained at the upstream portion of the watershed that went undetected by the rain gauge. We tried different climate data sources to improve model performance. However, current rainfall data offered by the USGS station proved to be the best data source. Two other alternatives to the USGS gauge was a NOAA rain gauge and NEXRAD radar. USGS rain gauge is at the watershed outlet. The NOAA rain gauge is about 16 km away from the Magnolia River watershed outlet and well outside its boundaries. Further, it records daily rainfall from 6:00am to 6:00pm, thus does not represent a calendar day. This may cause problems in daily flow simulations if there is an overnight rain event. The NOAA rain gauge also had extended periods of missing data (e.g. the whole months of November 2002, December 2002, and September 2009 were missing). Like the USGS rain gauge data we observed inconsistencies during big rainfall events in NOAA data. Summer rains in Alabama are dominantly localized pop-up thunderstorms. Capturing these storms requires a very dense network of rain gauges. Radar data seems to be a good alternative but that has its own problems too. We obtained NEXRAD radar data for the Magnolia River watershed for the 2002-2008 period and tried to calibrate the model. Even NEXRAD data did not capture rainfall accurately and we had poorer model performance. The 14 annual average precipitation from 2002 to 2008 based on NOAA rain gauge, NEXRAD and USGS rain gauge were 1794 mm, 1520 mm and 1315mm, respectively, which shows the discrepancies between these three rainfall datasets and the degree of spatial variation in this area. Once the models were calibrated for flow, they were calibrated subsequently for sediment and nutrients. Since measured sediment and water quality data are discontinuous, USGS?s LOADEST (A Fortran Program for Estimating Constituent Loads in Streams and Rivers) is applied to generate continuous loads when given a time series of streamflow, additional data variables and constituent concentration based on regression analysis. By LOADEST, a continuous monthly loadings of TSS and nutrient are generated as observed data. Due to lack of sufficient measured water quality data, monthly sediment and nutrient was calibrated for year 2000 and validated for year 2001. Model performances are shown in Table 4 and Fig. 3. Calibrated model parameters along with their default values are shown in Table 5 From Fig 3, we found SWAT is able to predict the monthly sediment and nutrient loadings with sufficient accuracy. Transferability of model parameters from Magnolia River to Wolf Bay watershed In the previous section, SWAT was manually calibrated for flow in Magnolia River watershed with the calibrated parameters shown in Table 2. The next step is transferring these parameters systematically to Wolf Bay watershed. Table 6 shows the daily model performances with the default and transferred parameter sets. Although SWAT performed better with the transferred parameters, ENS is negative and mass balance error is above 50%. Default parameters resulted in a much lower ENS value (compare -2.07 to -0.21). Although, to some extent we expected low performance with the default parameters, having such low performance with the 15 transferred parameters was surprising since watersheds have similar physical and morphological characteristics and are adjacent to each other. Note that in spite of low ENS, R2 is high. High R2 with a low ENS means simulated values have the same trend with observed values in time, but at a disproportionate rate. In other words the model systematically over/under predicts the observed data. In this case, it is an over prediction. Sometimes when models are run outside their calibration/validation periods, changing LULC may result in poor model performance. Model was set up using LULC map of 2005, but the simulation period was extended to 2009. If LULC changed significantly from 2005 to 2009, then model performance should deteriorate over time. However, LULC didn?t change much in that period. We also run SWAT with LULC map of 2008 and compared the daily simulation results to the ones obtained with 2005 LULC. No significant difference was detected between the two daily simulation results. Based on above findings it is seen that parameter transferring improves the predictable capabilities of the SWAT model in the study area, but not necessarily at the desired level. However, we calibrated and validated the model over a long time period (5+5 years) and tested the model with transferred parameters over a short period (~2 years). Whether the model performs well at the donor watershed during the testing period 10/2007 - 09/2009 is not clear. Note that although the validation period included this period, the length of the validation period (5 years) along with higher flows during the first year (2005) could potentially hinder the model performance in the testing period. If the model cannot accurately predict flow in this specific period at the donor watershed, then the problem is beyond parameter transferability. Indeed, model performance in Magnolia River watershed during this testing period is not good at all. R2, ENS and MBE for daily streamflow are 0.45, -0.21, and 48%, respectively, quite similar to what 16 we attained in Wolf Bay watershed with the transferred parameters. Exchanging the roles of donor and target watersheds, that is calibrating the SWAT model at Wolf Bay watershed and transferring the calibrated model parameters to Magnolia River watershed, resulted in a different story. Daily model performance statistics for the period 10/2007 - 09/2009 was R2=0.63, ENS= 0.62, and MBE=-2.5% at the calibration and R2=0.54, ENS= 0.51, and MBE=12% at the test watersheds. This is a substantial improvement over the previously reported values based calibration at Magnolia River watershed. Therefore, transferring parameters from neighboring watersheds indeed improves the predictive power of the model. Effect of parameter transferring on hydrologic responses In previous section we showed that transferred parameters increase model reliability as opposed to using model default parameters. Here we explore the implications of this on hydrological responses. Monthly flow simulation results using default and transferred parameters are shown in Fig. 4. Visually there are no significant differences. The two-sample Kolmogorov? Smirnov (K-S) test, which is sensitive to differences in both location and shape of the empirical cumulative distribution functions of the two samples, is employed to compare the two streamflow time series. The K-S test indicated no significant differences (p=0.482) in simulation results of monthly flow due to use of two separate parameter sets. However, at the daily time scale there are striking differences. As shown in Table 7, two specific years, 2000 and 2005 are selected to represent dry and wet conditions, respectively. The K-S test is employed to check if there are any significant changes in the time series obtained by the two parameter sets, both at daily and monthly time scales. Again, no significant differences exist at monthly time scale, both 17 in dry and wet years (p=0.518 and 0.848, respectively). However, if the simulation scale is changed to daily step, significant differences appear in both years (p<0.0001 in both). FDCs for these two years and the whole 10-year-period were also compared (Fig. 5a-c). Differences are evident at high and low flows, regardless of dry or wet year. Note that simulations with default parameters always resulted in higher flows at low exceedance probabilities (<1%). Being in a wet or dry year did not change the fact that if default parameters are used in predicting high-flows we will end up over predicting the flow. Similarly, default parameter set consistently generated lower flows than transferred parameter set when probability of exceedance was larger than 3% during both dry and wet years. Similar results were obtained when the FDC for the whole period was considered. Default parameters always resulted in higher flows at low probability of exceedance and lower flows at high probability of exceedance. Impact of LULC change on hydrologic responses Table 8 shows the LULC distributions in 1992 and 2005. From 1992 to 2005 percent forest cover has been reduced by 9%. On the contrary, total urban land has increased by almost 20%. Pasture has been lost to agricultural fields such as sod farming, and low density residential areas. Same climate data and parameter set (the one calibrated in Magnolia River watershed) were utilized as model inputs to run SWAT with both 1992 and 2005 LULC maps. Simulation results for flow for each year are summarized in table 9. For each year, we tabulated annual maximum, minimum and mean daily flow values obtained with LULC of 1992 and 2005. The relative change in mean annual flow due to LULC change shows little variation. Although change in LULC did not have a big impact on streamflow, it affected the partitioning of streamflow to baseflow and surface runoff as evidenced by changes in annual maximum and minimum flows. 18 In every single year annual minimum flow was predicted to decrease due to LULC change with a range from -16.8% to -36.9%, and an average of -29.7%. Annual maximum flow appears to be most sensitive to LULC changes. It is estimated to increase by 40.6% to 115.3% with an average of 58.0%. Similar results are found in other studies (e.g. Kauffman et al. 2009, Rose and Peter 2001). Fig. 6 shows the variation in average monthly streamflows before and after LULC changes. Flow was predicted to increase as much as 12% during the summer months of June through September, which is the growth season with the highest ET rates. Note that most of the increase in urban land was in form of low density residential areas (Table 8), which are only partially covered by impervious land. SWAT assumes that parts of urban areas not covered by impervious surfaces are bermudagrass. Based on SWAT database, maximum LAI for forest is 5, while for bermudagrass it is 4. Thus, forest to grass conversion does not cause significant change in ET. Over the whole simulation period, SWAT predicted about 20% less ET from grassland compared to forested land. Estimated annual average ET over the whole watershed based on the 1992 and 2005 LULC were 457 and 435 mm, respectively (~ 5% reduction). Low and high density residential areas have on average 12% and 60% imperviousness, respectively, according to the USDA Soil Conservation Service (SCS) classification. Thus, the total increase in percentage of impervious areas from 1992 to 2005 is only 2.95%. Since LULC from 1992 to 2005 did not change uniformly over the whole watershed, it is not reasonable to try to explain the alterations in flow and ET purely by changes in forest cover and urban land use. Note that from 1992 to 2005 pasture had also decreased by 28.3% and agricultural land increased by 15.5%. Therefore, there is a compound effect of all these mixed LULC changes. 19 Impact of parameter transferring on modeling LULC change Fig. 7a-c show the FDCs of daily flow simulations in the Wolf Bay watershed obtained with each of the three parameter sets (default, transferred from Magnolia River watershed, and the calibrated set in Wolf Bay watershed) and with the 1992 and 2005 LULC maps. In each figure there are clear differences in FDCs generated from the two LULC maps. Although the scales differ, FDCs with default parameters look similar to FDCs with transferred parameters (Fig. 7a- b). In both cases, the flow with LULC 2005 is higher than the one with LULC 1992 in the exceedance probability below 10%, and the relationship is opposite above 10%.When the calibrated parameters are used, flows based on 1992 LULC never seem to exceed flow based on 2005 LULC (Fig. 7c). To get a better insight into the effect of parameterization on the differences in FDCs due to LULC changes, relative difference in FDC, i.e. ([FDC of 2005 LULC] ? [FDC of 1992 LULC]) / [FDC of 1992 LULC], were depicted in Fig. 8. Trends are similar in all three. Moving from left to right, i.e. from low to high probability of exceedance, the relative differences in flow due to LULC change (from 1992 to 2005 conditions) increase until around 1% and stays at that level until 2%. Thus LULC change has the largest impact on flows with 1% exceedance probability. As flow gets larger (probability of exceedance < 1%) the impact of LULC change is gradually reduced. Beyond 2% all three parameter sets exhibit a sharp drop and reach a plateau again. With the default and transferred parameter sets, the relative change in flow becomes negative around 7-10% exceedance probabilities and stays negative beyond that point, mostly in the -15% to - 20% range. On the contrary, with the calibrated parameter set relative change in flow becomes negative around 20% exceedance probability and stays negative until 60% exceedance probability. Except for a short duration, the relative change in flow during this period is around 20 2%. Beyond 60% exceedance probability all the parameter sets show increase in relative change. In short, Fig. 8 reveals very interesting facts. First, the choice of the parameter set in simulating LULC changes does not seem to play a big role in relative changes of flow (it does for absolute flows). Although there are differences in the FDCs, trends are mostly consistent and the differences between them are not major. Secondly, LULC change influences most of the flow in a quite steady and moderate manner. Flows with 1-2% probability of exceedance appear to be most sensitive to LULC changes. SUMMARY AND CONCLUSIONS In this paper we explored how transferring model parameters from a neighbor (donor) watershed to a target watershed affects modeling LULC changes. The regionalization based on spatial proximity method was employed to transfer the model parameters from the donor Magnolia River watershed to the target Wolf Bay watershed. For this purpose SWAT model was first set up and calibrated in the Magnolia River watershed which has 10 years of continuous measured flow data. Calibrated model parameters were then transferred to the Wolf Bay watershed, which at the start of study had no flow data. Model performances were compared when two different parameter sets are utilized: SWAT default parameters and transferred parameters. About 22 months of measured flow data in Wolf Bay watershed that later became available was used for that purpose. Transferred parameter set resulted in a slightly better model performance than the default parameter set, but not at a desired level (both had negative ENS). The low model performance was due to the fact that when using a long period of data in model calibration the emphasis is on the whole period and the model performance may not be up to the desired level in some subsections of the entire time period. Hence, extension of parameter 21 transfer from donor watersheds through long term calibration to target watersheds for short term predictions requires extra caution, and most likely vice versa. Flow duration curves (FDCs) are effective tools in visualizing the whole flow range. We created FDCs to get a better insight to the impacts of both LULC change and parameter transferring. Simulations with default parameters always resulted in higher flows at low probability of exceedance (<1%). On the contrary, default parameter set consistently generated lower flows than the transferred parameter set when probability of exceedance was > 3%, regardless of dry or wet year. Similar results were obtained when the FDC for the whole period is considered. Default parameters always resulted in higher flows at low probability of exceedance, and lower flows at high probability of exceedance. Two LULC maps from 1992 and 2005 were utilized to assess the effect of parameter transferring on modeling LULC changes. The 2005 LULC had about 20% more urban classified land than the 1992 LULC. However, the estimated change in impervious cover from 1992 to 2005 was only 2.95%. Average streamflow was only slightly affected by LULC changes. Maximum and minimum annual streamflows were found to be very sensitive to LULC changes. Annual minimum streamflow decreased moderately and annual maximum streamflow increased substantially due to LULC change. Again FDCs were developed out of model generated daily flows based on 1992 and 2005 LULC maps. This was done for each of the three parameter sets: default, transferred and Wolf Bay calibrated. The relative changes in FDCs due to differing LULC showed a similar pattern with each parameter set: relative change was highest at 1-2% exceedance probability. The impact of LULC change diminished gradually as the event sizes got smaller beyond the 2% probability of exceedance. This study clearly showed the benefits of using data from nearby watersheds to improve the model performance under limited data conditions in the study watershed. The analysis carried out in this study further suggest that t parameters or transferred from a donor watershed, only has a marginal effect on modeling the impacts of different LULC scenarios. Arnold JG, Allen PM, Muttiah R, Bernhardt G. 1995. Au Recession Analysis Techniques. Arnold JG, Srinivasan R, Muttiah RS, Williams JR. 1998. Large Area Hydrologic Modeling and Assessment - Part 1:Model Development. Association (JAWRA) 34(1): 73 ASCE. 1993. Criteria for evaluation of watershed models. Engineering 119(3): 429-442. Baffaut C, Benson VW. 2009. Modeling Flow and Pollutant Transport in a Karst Watershed with SWAT. Transactions of the ASABE Chanasyk, DS, Mapfumo E, Willms W. 2003. Quantification and Simulation of Surface Runoff From Fescue Grassland Watersheds. El-Nasr AA, Arnold JG, Feyen J, Using a Distributed and a Semi Fohrer N. 2002. River basin research and management: influence of land cover, land use and soil surface conditions on hydrological processes and water balance of catchment. Chemistry of the Earth. 27: 617. 22 he choice of the parameter set, whether it is default model REFERENCES tomated Baseflow Separation and Ground Water 33(6): 1010-1018. Journal of the American Water Resources -89 Journal of Irrigation and Drainage 52(2): 469-479. Agricultural Water Management 59(2): 137 Berlamont J. 2005. Modelling the Hydrology of a Catchment -Distributed Model. Hydrological Processes 19(3) -153. :573-587. Physical and Fohrer N, Haverkamp S, Eckhardt K, Frede HG. 2001. Hydrologic Response to Land Use Changes on the Catchment Scale. Oceans and Atmosphere 26: 577 Guo H, Hu Q, Jiang T. 2008. Annual and seasonal streamflow responses to climate and land cover changes in the Poyang lake basin, China. Hanratty MP, Stefan HG. 1998. Simulating Climate Change Effects in a Minnesota Agricultural Watershed. Journal of Environmental Quality Hundecha Y, Bardossy A. 2004. Modeling of the effect of land use changes on the runoff generation of a river basin through parameter regionalization of a watershed model. of Hydrology 292(1-4): 281-295 Hevnelmans G, Muys B, Feyen J. 2004. Evaluation of hydrological model parameter transferability for simulating the impact of land use on catchment hydrology. Chemistry of the Earth 29: 739 Kalin L, Hantush MM. 2006. Hydrologic Modeling of an Eastern Pennsylnania Watershed with NEXRAD and Rain Gauge Data. Kauffman GJ, Belden AC, Vonck, KJ, Homs base flow in the White Clay Creeek Wild and Scenic watershed in Delaware. Hydrologic Engineering. 14(4) Kim U, Kaluarachchi J. 2008. Application of parameter estimation and regionalizati methodologies to ungauged basins of the Upper Nile River Basin, Ehiopia. Hydrology 362(1): 39-56. 23 Physical and Chemistry of the Earth, Part B: Hydrology, -582. Journal of Hydrology 355(1-4) 27(6): 1524-1532. . -747. Journal of Hydrologic Engineering 11(6): 555 ey AR. 2009. Link between impervious cover and : 324-334. Journal of - : 106-122. Journal Physics and -569. Journal of on 24 Larose M, Heathman GC, Norton DL, Engle B. 2007. Hydrologic and atrazine simulation in the cedar creek watershed using the SWAT model. Journal of Environmental Quality 36(2):521- 531. Lee D, Kim N, Chung I, 2005. Comparison of runoff responses between SWAT and sequentially coupled SWAT-MODFLOW model. International SWAT 2005 Conference, Zurich, Switzerland. http://www.brc.tamus.edu/swat/3rdswatconf/PDF/Session_VII/Lee.pdf. Last access on 07/01/2010. Legates DR, McCabe GJ. 1999. Evaluating the use of ?goodness-of-fit? measures in hydrologic and hydro climatic model validation. Water Resources Research 35(1):233-241. Li Z, Liu W, Zhang X, Zeng F. 2009. Impacts of land use change and climate variability on hydrology in an agricultural catchment of the loess Plateau of China. Journal of Hydrology 337(1-2):35-42. Ma X, Xu J, Luo Y, Aggarwal SP, Li J. 2009. Response of hydrological processes to land-cover and climate changes in Kejie watershed, South-West China. Hydrological Process 23(8): 1179-1191. Merz R, Bloschl G. 2004. Regionalisation of catchment model parameters. Journal of Hydrology 287(1-4): 95-123. Moriasi DN, Arnold JG, Van Liew MW, Bingner RL, Harmel RD, Veith TL. 2007. Model evaluation guidelines for systematic quantification of accuracy in watershed simulation. Transactions of the ASABE, 50(3): 885?900. Nash JE, Sutcliffe JE. 1970. River Flow Forecasting Through Conceptual Models, Part I-A Discussion of Principles. Journal of Hydrology 10(3):282-290. 25 Neitsch SL, Arnold JG, Kiniry JR, Williams JR. 2005. Soil and Water Assessment Tool: Theoretical Documentation, version 2005. http://www.brc.tamus.edu/swat/downloads/doc.html. Last accessed on 07/01/2010. Oudin L, Andreassian V, Perrin C, Michel C, Moine NL. 2008. Spatial proximity, physical similarity, regression and ungaged catchments: A comparison of regionalization approaches based on 913 French catchments. Water Resources Research 44, W03413. Reichl PC, Western AW, McIntyre NR, and Chiew HS. 2009. Optimization of a similarity measure for estimating ungauged streamflow. Water Resources Research 45, W10423. Rose S, Peters NE. 2001. Effects of urbanization on streamflow in the Atlanta area (Georgia, USA): a comparative hydrological approach. Hydrological Processes 15(8): 1441?1457 Santhi C, Arnold JG, Willians JR, Dugas WA, Srinivasan R, Hauck LM. 2001. Validation of the SWAT model on a large river basin with point and nonpoint sources. Journal of the American Water Resources Association (JAWRA) 37(5): 1169-1188. Savary S, Rousseau AN, Quilbe R. 2009. Assessing the Effects of Historical Land Cover Changes on Runoff and Low Flows Using Remote Sensing and Hydrological Modeling. Journal of Hydrologic Engineering 14(6): 575-587. Sorooshian S, Gupta VK, Fulton J L. 1983. Evaluation of maximum likelihood parameter estimation techniques for conceptual rainfall-runoff models: Influence of calibration data variability and length on model credibility. Water Resource Research 19(1): 251?259. Stallman C., McIlwain K, Lemoine D. 2005. The Baldwin County Wetland conservation Plan Final Summary Document. Http://www.wetlands.co.baldwin.al.us. Last accessed on 07/01/2010. 26 Tu J. 2009. Combined impact of climate and land use changes n streanflow and water quality in eastern Massachusetts, USA. Journal of Hydrology 379(3-4): 268-283. Wang X, Melesse AM. 2005. Evaluation of the SWAT Model?s Snowmelt Hydrology in a Northwestern Minnesota Watershed. Transactions of the ASAE 48(4):1359-1376. Wu K, Xu Y. 2006. Evaluation of the applicability of the SWAT model for coastal watersheds in southeastern Louisiana. Journal of the American Water Resources Association (JAWRA) 42(5): 1247-1260. Xia Y, Yang Z, Jackson C, Stoffa PL, Sen MK. 2004. Impacts of data length on optimal parameter and uncertainty estimation of a land surface mode. Journal of Geophysical Research, 109: 142-157. Yokoo Y, Kazama S, Sawamoto M, Nishimura H. 2001. Regionalization of lumped water balance model parameters based on multiple regression. Journal of Hydrology 246(1-4): 209- 222. 27 Table 1. Physical similarities between Wolf Bay and Magnolia River watersheds Physical characters Wolf Bay Magnolia Min elevation (m) 0 6 Max elevation (m) 34 36 Mean elevation (m) 16.65 25.31 Area (km2) 126.04 44.82 Rural area 2005 (%) 72.8 73.65 Urban area 2005 (%) 27.2 26.35 Soil Clay (%) 8.61 12.33 Soil Silt (%) 18.40 24.24 Soil Sand (%) 72.99 63.43 Mean slope 1.88 1.42 Table 2. Calibrated SWAT parameters (flow part) and their default values Curve Soil ESCO surlag revapmn Alpha_BF Manning's Number AWC n Default Varies* Varies** 0.95 4 10 0.048 0.014 Calibrated 3 -0.01 1 1 500 0.015 0.114 * Varies by soil type and LULC ** Varies by soil type 28 Table 3. Model performance (flow part) at Magnolia River watershed R2 ENS MBE (%) Monthly streamflow Calibration 0.84 0.82 -7.1 Validation 0.80/0.78* 0.65/0.74* -2.0/4.7* Monthly surface runoff Calibration 0.88 0.83 3.6 Validation 0.83 0.68 -4.8 Monthly baseflow Calibration - - -13.4 Validation - - 0.4 Daily streamflow Calibration 0.51 0.50 -7.1 Validation 0.45/0.54* 0.39/0.54* -2.0/4.7* * Model performance after removing the extreme event on 04/01/2005 Table 4. Model performance (water quality part) at Magnolia River watershed R2 Ens MBE (%) Monthly TSS Calibration 0.90 0.85 8.7 Validation 0.93 0.88 -2.2 Monthly Min-P Calibration 0.95 0.86 9.4 Validation 0.87 0.77 -21.3 Monthly Org-P Calibration 0.97 0.95 15.7 Validation 0.78 0.76 -1.8 Monthly Org-N Calibration 0.93 0.92 -11.6 Validation 0.66 0.61 -7.4 Monthly Min-N Calibration 0.74 0.60 -14.8 Validation 0.87 0.85 4.7 Monthly TP Calibration 0.96 0.89 11.0 Validation 0.86 0.80 -16.7 Monthly TN Calibration 0.75 0.62 -14.5 Validation 0.88 0.86 4.5 TSS: Total suspended solid Min: Mineral nutrient Org: Organic nutrient 29 Table 5. Calibrated SWAT parameters (water quality part) and their default values Table 6. Model performance at daily time scale in Wolf Bay watershed See text for explanation of terms Table 7. Kolmogorov?Smirnov test for daily and monthly simulated streamflows generated with different parameter sets. Kolmogorov?Smirnov test Daily average rainfall (mm) Daily Max rainfall (mm) year KSa daily p-daily KSa monthly p-monthly 2000 (dry) 11.421 <0.0001 0.816 0.518 1.1 192.8 2005 (wet) 2.924 <0.0001 0.612 0.848 4.8 50.5 The significance of KSa is 0.95 BC4 PSP PHOSKD BC1 PPERCO RS5 AGRRC Sol_minP Default 0.35 0.4 175 0.55 10 0.05 0.3 5 Calibrated 0.1 0.7 200 1 17.5 0.1 0.055 3 PRF BC3 P_UPDIS BC2 NPERCO RS4 RCN Mgt for AGRR Default 1 0.21 20 1.1 0.2 0.05 1 Auto fertilize, heat unit Calibrated 0.6 0.4 100 2 1 0.001 2 Cotton peanut rotation, date Default Transferred Calibrated Parameters Parameters Parameters MBE(%) 0.478 0.516 -0.025 R2 0.536 0.637 0.63 ENS -2.067 -0.208 0.618 30 Table 8. Land use/cover (LULC) change in Wolf Bay watershed Land Use 1992(%) 2005(%) Change(%) Forest 29.80 20.70 -9.10 Hay 41.20 12.90 -28.30 Wetland 11.20 13.40 2.20 Agricultural 13.30 28.80 15.50 Residential low density 3.50 21.60 18.10 Residential high density 1.10 2.40 1.30 Table 9. Annual statistics of streamflow under different LULC conditions Year Maximum flow (m3s-1) Minimum flow (m3s-1) Mean flow (m3s-1) 1992 2005 Change(%) 1992 2005 Change(%) 1992 2005 Change(%) 1999 7.26 11.74 61.7 0.266 0.194 -27.2 0.835 0.793 -5.01 2000 1.15 2.48 115.3 0.071 0.054 -24.6 0.175 0.193 10.08 2001 1.30 2.67 104.7 0.127 0.105 -16.8 0.255 0.253 -0.65 2002 14.25 21.67 52.1 0.078 0.062 -20.1 0.428 0.509 18.80 2003 8.53 13.63 59.8 0.317 0.243 -23.3 1.003 1.036 3.32 2004 11.30 17.66 56.3 0.438 0.276 -36.9 0.876 0.901 2.85 2005 18.09 25.43 40.6 0.529 0.411 -22.4 1.247 1.213 -2.78 2006 2.81 5.48 95.0 0.196 0.138 -29.3 0.458 0.481 5.19 2007 8.40 13.70 63.1 0.250 0.175 -29.9 0.592 0.620 4.63 2008 4.21 6.94 64.9 0.515 0.331 -35.8 0.855 0.866 1.33 2009 4.84 8.41 73.8 0.510 0.327 -35.9 0.799 0.810 1.38 Mean 7.47 11.80 58.0 0.300 0.211 -29.7 0.684 0.698 1.98 31 Fig.1 Geographical location of Wolf Bay and Magnolia River watersheds 32 Fig.2 Simulated streamflow compared with observed data from 1999 to 2009 in Magnolia river watershed , (a) monthly time series (b) daily flow duration curve 33 Fig.3 Simulated monthly sediment and nutrient compared with observed data (LOADEST) from 2000 to 2001 in Magnolia River watershed 34 Fig.4 Monthly simulation streamflow for different parameter sets in Wolf Bay watershed 35 Fig.5 Flow duration curve using different parameter sets in Wolf Bay watershed, (a) dry year 2000, (b) wet year 2005, (c)whole period 1999-2009 36 Fig.6 Average monthly flow before and after land use change 37 Fig.7 Flow duration curve using different land use maps in Wolf Bay watershed by (a)default (b)transferred (c)calibrated parameters 38 Fig.8 Relative change in simulated streamflows due to different LULC for different parameters in Wolf Bay watershed (1999-2009) 39 Chapter II Responses of Hydrological Processes and Water Quality to LULC and Climate Change in Wolf Bay Watershed, Southern Alabama ABSTRACT Land use/cover (LULC) and climate change are two main factors affecting watershed hydrology and water quality. In this chapter, the individual and combined impacts of LULC and climate change on flow and water quality were analyzed by SWAT model by simulating the future changes under different LULC and climate change scenarios in the Wolf Bay watershed. Global Circulation Models (GCM) predict slight increase in precipitation in the Wolf Bay watershed, which is projected to experience substantial increase in urban percentage in the future. A redistribution of daily streamflow is projected when either climate or LULC change was considered. High flows are predicted to increase, while low flows are expected to decrease. Combined change effect results in more noticeable uneven distribution of daily streamflow. Monthly average streamflow and surface runoff are projected to increase in spring and winter, but especially in fall, under normal future climate conditions. LULC change does not have a significant effect on monthly average streamflow, but affect partitioning of streamflow, causing higher surface runoff and lower baseflow. Combined effect led to more dramatic increasing trend in monthly average streamflow with a stronger increasing trend in surface runoff and decreasing trend in baseflow. Monthly distribution and projected variation of TSS followed the pattern of flow. Monthly distribution and projected variation of nutrients are complicated, which are 40 influenced by flow as well as management practices, such as tillage, fertilization and harvesting. Under the climate change scenarios, the variation of TN and TP generally followed the trend of flow. No significant difference in N:P ratio was projected. Under the LULC change scenarios, TN was projected to decrease for all months, which is induced by shrinkage of croplands. TP was projected to increase in fall and winter, since urban areas are also source for TP. The N:P ratio showed a strong decreasing trend with LULC changes. Under the combined change scenario, LULC and climate change effect were considered simultaneously. Results indicate that if future loadings are expected to increase/decrease under either climate or LULC change scenarios, combined change scenario intensifies that trend synergistically. On the other hand, if their effects are in opposite directions, then the combined change has an offsetting effect. INTRODUCTION Alteration in flow regimes and water quality deterioration due to land use/cover (LULC) and climate change are of great concern all over the world. LULC changes, mostly caused by human activities including changes in vegetation types, soil properties, land use practices and spatial patterns of interactions among these factors, affect water quantity and quality, often negatively. Many studies have been conducted to explore this strong influence. Zhang and Schilling (2006) found that conversion of perennial vegetation to seasonal row crops in the Mississippi River basin has partly contributed to the increasing trend of baseflow and streamflow. Ouyang et al. (2010) studied soil erosion dynamics in response to landscape pattern and found that landscape pattern plays an important role in soil erosion. For instance, smaller patch size and more patch edge led to lower sediment loads in grasslands. 41 The most commonly observed LULC change is due to urbanization, which has been intensively studied in recent years. Urbanization leads to an increase in impervious areas which decrease the amount of water that infiltrates into the soil. Thus, while baseflow contribution to streamflow reduces, runoff increases, which results in more frequent and intense flooding (Rose and Peter, 2001; Huang et al., 2008). Urbanization also affects water quality adversely. It causes increase in sediment and nutrient loads, heavy metals, blooming of toxic algae which can reduce dissolved oxygen levels in waters (Bakri et al. 2008; Susana et al. 2008). Kim et al. (2002) modeled the changes in average annual runoff due to urbanization in the Indian River Lagoon Watershed of Florida. They found that the average annual runoff increased by more than 113% between 1920 and 1990. Ouyang et al. (2006) assessed the impact of urbanization on river water quality in the Peral River delta zone, China. They found that urbanization and urban activities had a significant negative impact on the river water quality with significant increase in nutrient loadings and turbidity. Studies show that climate change leads to intensification of the global hydrological cycle and has a major impact on regional water resources, which affects both the distribution and availability of water resources and in turn influences processes controlling water quality, such as erosion, sediment transport and deposition, settling of nutrient and pollutions (Dam, 1999; Oki and Kanae et al., 2006; Konikow and Kendy, 2005). Not only deterioration in water quality is a problem by itself, but also it contributes to the problem of water scarcity. Ficklin et al. (2009) assessed the climate change effect on San Joaquin Valley watershed in California and found that under future scenarios, streamflow will increase by 23.5%. Marshall and Randhir (2008) investigated the effect of climate change in the Connecticut River watershed, New England, by employing two downscaled GCM model outputs. They found that due to warming in climate 42 water storage will decrease during the winter months. They further predicted increased sediment loadings in summer months in spite of a decline in surface runoff rate. This was because of antecedent moisture conditions, variability in sediment transport capacity resulting from soil characteristics, and detachment process caused by higher precipitation. N:P ratio was projected to increase, resulting in the watershed becoming more nitrogen limited. Cruise et al. (1999) coupled the United Kingdom Hadley Center climate model with a regional stochastic approach and a physically based soil moisture model in the southeastern U.S. Results of their study revealed that several basins located in regions of intense agricultural activity or in proximity to the gulf coast are projected to have reduction in streamflow over the next 30-50 years, thus exacerbating water quality problems, such as high nitrogen concentration levels. From the past studies it is obvious that both LULC and climate change play key roles for water resources and water quality, yet their combined effect and relative importance is still not very clear, difficult to separate empirically, and varies from case to case. Several studies (Qi et al., 2009; Liu et al., 2010; D?Agostino et al., 2010) explored the combined effect from both LULC and climate change. Mango et al. (2010) used modeling to determine the impacts of LULC and climate change scenarios on the water flux of the upper Mara River, Africa. They found that deforestation resulted in a slightly more erratic discharge while rainfall and air temperature changes had a more predictable impact on the discharge and water balance components. Guo et al. (2008) studied the combined effects of climate and LULC change to hydrological processes in Poyan Lake basin, China. They found that climate change is more likely to alter annual streamflow, while LULC change may have a moderate impact. Both of them strongly influenced seasonal variation in streamflow. Olivera and Defee (2007) studied urbanization effects on a 223 km2 watershed located in the northwest suburbs of Houston, TX. 43 They found that runoff depths and peak flows have increased by 146% and 159%, respectively, since early 1970s. Urbanization was responsible for 77% and 32% of the increase, respectively, while variation in precipitation accounted for the remaining increase of 69% and 127%. As a valuable tool for studying the processes governing impacts of climate and LULC change on water quantity and quality, modeling is an inherently probabilistic exercise (Praskievicz and Chang, 2009). Generally, there are three different aspects of uncertainty in assessing impacts on hydrology and water quality. The first source of uncertainty is from future climate. This could come from (i) choice of the Global Circulation Models (GCMs) and future greenhouse gas emissions scenarios, and (ii) representation of climatology at regional scales, including differences between dynamical and statistical downscaling methods. The second source is associated with future LULC conditions, which are quite hard to predict and are often affected by land use policy, economic development, population increasing rate and natural environment. The third source of uncertainty stems from hydrologic models, such as model parameter estimation and model structure, like mathematical representation of the physical processes, which require many assumptions and simplifications. When projecting the impacts from potential future changes, uncertainty is inevitable and amplifies at each stage of the modeling process. Therefore, addressing uncertainties stemming from modeling potential LULC and climate change impacts and their combined effect is essential. Although there are several studies in the literature focusing on the combined effects of LULC and climate change on water quantity/quality, the following research gaps remains still exist: 1. Most of the previous studies are confined in the aspect of water quantity (Ma et al., 2009; Li et al., 2009; Cuo et al., 2009), with few studies addressing the effect on both water quality and quantity. Those studies concerned with the combined effects of climate and LULC change on 44 sediment or nutrients involved very limited water quality indices. For example, Tu (2010) assessed the seasonal distribution and annual averages of nitrogen loadings in eastern Massachusetts, USA. Similarly, Asselman et al. (2003) estimated the potential effects of climate combined with LULC changes on the mobilization of fine sediment and the net transport of wash load from the upstream basin to the lower Rhine delta. However, influences on nutrient were not stated in their paper. 2. Compared to future climate change scenarios, which usually contain various GCM outputs under different green house gas emission scenarios, LULC change scenarios are too simplistic and do not consider the factors affecting LULC changes, such as land use policy, economic development, and natural environment. For example, Montenegro and Ragab (2010) explored the hydrological response of a Brazilian semi-arid catchment to combined effect of LULC and climate change. However, their land use change scenario was hypothetical and overly simplified, replacing different amounts of catinga forest with castor beans. 3. The uncertainties of model results are not satisfactorily addressed in most of the studies. As stated before, the uncertainty in the model output originates from many sources. Some studies only acknowledged the uncertainty caused by climate inputs. For example, Wilby et al. (2006) studied climate change impacts on water resources and water quality in a British lowland catchment by 3 GCMs and 2 green house emission scenarios. Their results confirmed the large uncertainty in climate change scenarios and freshwater impacts due to the choice of GCM. However, there are limited studies dealing with the uncertainties from climate combined with LULC change. In this study, a process-based hydrologic and water quality model, Soil and Water Assessment Tool (SWAT) was utilized to simulate flow, sediment and nutrient (N and P) 45 loadings under future climate and LULC conditions in Wolf Bay watershed, which drains to the Gulf of Mexico. The specific objectives are: (1) to explore the hydrologic and water quality responses to combined effects of climate and LULC change, (2) to examine whether climate change exacerbates or offsets the impacts of LULC change and vice versa. Uncertainties caused by climate and LULC change are also analyzed for both objectives. METHODOLOGY Study Area Wolf Bay and its watershed (Fig. 1) is located on the Gulf of Mexico in Baldwin County, Alabama, nestled between Pensacola Bay to the east and Mobile Bay to the west, with a watershed covering about 126 km2. As an estuary where freshwater and saltwater mix, it creates a diverse environment that fosters a rich array of plant and animal life, including several federally listed species, such as bald eagles, sea turtles, Gulf sturgeons, American alligators and Eastern indigo snakes. Wolf Bay and its surrounding waters are the most pristine estuarine waters in Alabama, which was granted ?Outstanding Alabama water? status by Alabama Department of Environmental management in April, 2007. The beautiful waters attract many people to coastal Baldwin County, contributing greatly to the economic base of coastal communities through tourism, commercial and recreational fishing and aquaculture. Wolf Bay watershed is primarily rural, but several municipalities exist including Foley, Elberta, Gulf Shores, and Orange Beach (Fig. 1). The basin generates various nutrient and sediment inputs from several sub-watersheds through Wolf, Sandy, Miflin and Hammock creeks, which finally drain to the Wolf Bay. Alteration in vegetations, management practices of the watershed 46 can change hydrology and water quality, and in turn can significantly affect the water resource and ecologic health of the bay. Baldwin County experienced a 43% increase in population from 1990 to 2000. As a result of this population growth, there has been an increased demand for commercial, residential, and infrastructure development, thus bringing growth management issues to the forefront for local elected officials. One of the more visible changes in the landscape of Baldwin County is the rapid transformation of agricultural and forested lands to residential development. Such LULC change is deemed to affect water quantity and quality, usually negatively. Considering the potential climate change effects, the situation becomes more complicated. Their potential effects can be reflected in: (1) Water quality: Increased soil erosion can change the shoreline from sandy to muddy, which could destroy the fish stock and damage the benthos and habitats. Increased turbidity also degrades the ecosystem by decreasing the light available for photosynthesis. Further, excessive nutrients can lead to water quality degradation, which reduce dissolved oxygen, causing hypoxia or anoxia. This may destruct the whole ecosystem by blooming harmful algae and therefore causing massive fish kills. (2) Water quantity: Due to climate and urbanization, it is anticipated that peak flows will increase and dry season flows will decrease, thus exacerbating flooding in wet seasons and droughts in dry seasons. This will also affect water quality indirectly. For example, if the estuary is slowly flushed, the extra load of nutrients and pollutants will cause degradation in water quality, ecological services and biodiversity. Also, fluctuations in freshwater discharge to the bay affect the salinity of water, which affects the health and incubation of fishes. For example, a 47 slight change in salinity can cause fish, frog or shrimp eggs to float too much (high salinity) or not enough (low salinity), thus reducing or eliminating their chances of development into adults. Because of all these, we found a unique opportunity to study the potential impacts of LULC and climate changes on hydrologic responses and water quality in the Wolf Bay watershed. Findings from this study should benefit local stakeholders and decision makers in the Wolf Bay area. Watershed model The Soil and Water Assessment Tool (SWAT) version 2005 (Neitsch et.al, 2005) was used in this study. SWAT is one of the most widely used models for assessing the impact of management practices and land disturbances on watershed responses, and has a solid track record of applications (Kalin and Hantush, 2006 ; Gul and Rosbjerg, 2010; Pisinaras V et al., 2010). SWAT has been widely used around the world, such as Nzoia catchment, Kenya (Githui et al., 2009), Rocky Mountain Watershed, Montana, USA (Ahl et al., 2008), Kielstau catchment in North German lowlands (Lam et al.,2010), etc., to assess various impacts of agricultural practices and land use activities on water quantity and quality. SWAT is also suitable for coastal and flat areas, which has more complicated geo-hydrologic conditions (Wu and Xu, 2006). SWAT is a distributed, process-based watershed model. It is partly physical-based with number of empirical relationships. The physical backbone of the model facilitates the interpretation of model parameters whereas the empirical simplifications keep data requirements low compared to fully physical based models (Heuvelmans, 2004). SWAT divides a watershed into several subwatersheds based upon drainage areas of the tributaries. Each subwatershed is split into multiple hydrological response units (HRUs) based on LULC and soil types. Each HRU is 48 assumed to be spatially uniform in LULC, soil, topography, and climate. SWAT simulates eight major components: hydrology, weather, sediment, soil temperature, crop growth, nutrients, pesticides, and agricultural management (Neitsch et al., 2005). Major hydrologic process that can be simulated by the model include evapotranspiration, surface runoff, infiltration, percolation, shallow aquifer and deep aquifer flow, and channel routing (Arnold et al., 1998). Erosion and sediment yield are estimated for each HRU with the Modified Universal Soil Loss Equation (MUSLE) (Williams, 1975). Sediment routing is also considered based on deposition and degradation processes. SWAT also tracks the movement and transformation of several forms of nutrients (phosphorus and nitrogen) in the soil. Nutrient may be introduced to the main channel by surface or subsurface runoff, nutrient routing in the stream is then controlled by the in-stream water quality component adapted from QUAL2E (Brown and Barnwell, 1987). Detailed description of processes modeled in SWAT can be found in Neitsch et al. (2005). LULC data In order to explore the LULC change effect on hydrology and water quality in the Wolf Bay watershed, present and projected LULC maps are needed. LULC map circa 2005 is used to represent the current period. It is a vector dataset attained by trend analysis from Baldwin County Planning Commission. This vector dataset is focused on changes in urban and built-up areas, utilities, and transportation from 2001 to 2005 based on Color Infrared imagery 2001 and 2005 for the whole Baldwin County. Using this trend map as a reference, GIS specialists at Auburn University improved its accuracy and produced LULC map of 2005. This map is a product of an interdisciplinary project ?Impact of Human Activities and Climate Change on Water Resources and Ecosystem Health in Wolf Bay Basin: A Coastal Diagnostic and Forecast System (CDFS) 49 for Integrated Assessment?. Based on this map, Wolf Bay watershed is dominated by agricultural land (30%) followed by urban area (26.4%) and forest (20.9%). High percentage of urban area and crop land is due to conversion from forest and pasture land in the past 10-15 years. Future LULC of the Wolf Bay watershed was projected by members of the same interdisciplinary project at Auburn University. They developed an advanced LULC model by linking GIS techniques and remotely sensed images creating a hybrid model. The predicted LULC change is driven by land demands, physical properties such as topography and distance to major facilities, and disturbances such as extreme climate events (hurricanes, storms and droughts). The LULC prediction model can simply be described as: g1838g1847g1838g1829 g3404 g1858g4666g1864g1853g1866g1856 g1856g1857g1865g1853g1866g1856,g1871g1868g1853g1872g1861g1853g1864 g1853g1864g1864g1867g1855g1853g1872g1861g1867g1866,g1870g1853g1866g1856g1867g1865g4667 g1838g1853g1866g1856 g1856g1857g1865g1853g1866g1856 g3404 g1858g4666g1868g1867g1868g1873g1864g1853g1872g1861g1867g1866,g1857g1855g1867g1866g1867g1865g1877,g1864g1853g1866g1856 g1868g1867g1864g1861g1855g1877,g1857g1866g1857g1870g1859g1877 g1868g1870g1861g1855g1857g4667 g1845g1868g1853g1872g1861g1853g1864 g1853g1864g1864g1867g1855g1853g1872g1861g1867g1866 g3404 g1858g4666g1872g1867g1868g1867g1859g1870g1853g1868g1860g1877,g1864g1853g1866g1856 g1868g1870g1861g1855g1857,g1856g1861g1871g1872g1853g1866g1855g1857 g1872g1867 g1860g1861g1859g1860g1875g1853g1877g1871,g1871g1860g1867g1870g1857g1864g1861g1866g1857,g1870g1867g1853g1856g1871,g1870g1861g1874g1857g1870g1871,g1872g1860g1857 g1855g1861g1872g1877 g1868g1870g1867g1868g1857g1870g4667 g1844g1853g1866g1856g1867g1865 g3404 g1858g4666g1860g1873g1870g1870g1861g1855g1853g1866g1857,g1871g1872g1867g1870g1865,g1870g1861g1871g1861g1866g1859 g1871g1857g1853 g1864g1857g1874g1857g1864g4667 Based on this modeling framework, historical LULC data sets and the derived spatial data sets from DEM and survey data were used to generate preliminary simulation results on the projected urban distributions from 2008 to 2040. With the validated LULC model and relevant data sets, projected urban expansions with different population growth scenarios are provided. Three LULC scenarios (Fig. 2) were generated for 2030 assuming high, medium and low population increasing rates (HPR, MPR and LPR). Higher population increase rate causes higher urban fraction and vice versa. Compared with the most recent land use map of 2005, there is a clear trend of urban sprawl. Even with the least aggressive growth scenario, 50% of the watershed is projected to be urban land in 2030 (Table 1). Other LULC types are projected to decline by 2030 owing to the urbanization effect. For example, by comparing LULC map of 50 2005 with LPR future projected map of 2030, the evident increase of urban area (around 25%) is mainly contributed by decreases of forest, agriculture land, wetland and pasture. The percent reduction in forest cover is around 5%, which does not represent a typical deforestation trend in future. The disparity in some of the LULC types among the three projected LULC maps of 2030 is not so significant, especially for forest, pasture and wetland. The main difference is in percentages of urban and agricultural land. Higher population increase rate causes higher urban fraction and lower cropland percentage. For example, the increase of urban fraction is 25%, 32% and 39% for LPR, MPR and HPR, respectively while the reductions in cropland are 5%, 10% and 15%, respectively. Climate data Monthly precipitation and temperature for future scenarios In order to demonstrate the variability of future climate, outputs from four Global Circulation Models (GFDL_cm2_0 (Delworth et al., 2006), GISS_model_e_r (Russell et al., 2000), NCAR_ccsm3_0 (Collins et al., 2006), UKMO_hadcm3 (Gordon et al., 2000)) under 3 green house gas emission scenarios (A2, A1B, B1) were utilized to attain potential monthly precipitation and temperature estimates in the Wolf Bay watershed for the period 2016-2040. This corresponds to a 25 year period, which is long enough to explore the potential responses due to climate change. Further, the future LULC map of 2030 roughly falls in the middle of this time period, which presents a more realistic set up for exploring the combined effects of climate and LULC change. All climate projections were provided by "the World Climate Research Programme's (WCRP's) Coupled Model Intercomparison Project phase 3 (CMIP3) multi-model dataset" which 51 was referenced in the Intergovernmental Panel on Climate Change Fourth Assessment Report. CMIP3 data is bias-corrected and spatially downscaled by Maurer et al. (2007) to a finer spatial resolution (1/8 degree). Spatially downscaled monthly rainfall and surface air temperature data (available at: http://gdo-dcp.ucllnl.org/downscaled_cmip3_projections/) were further downscaled to daily time scale as explained below. Since there is no observed data for future scenarios, ensuring the reliability of GCMs projection is quite important. Generally, GCMs also provide simulated precipitation and temperature for historic periods, which should compare well with observed climate data. In most of the cases, those historic GCM outputs match well with observed data at large spatial scales (e.g. global or continent scale). However, once spatially downscaled, those products could become quite different from historic data at smaller scales, such as watershed level, which is often the required scale for hydrologic modeling. Therefore, even spatially downscaled climate projections cannot be directly utilized as climate input for hydrologic modeling. Refinement of those spatially downscaled data is often necessary. The method recommended by Tung et al. (2006) was used to determine monthly temperature and precipitation for different climate scenarios: g1846g3364g3038 g3404 g1846g3364g3038g3029 g3397 g3435g2028g1191g3038g3033 g3398 g2028g1191g3038g3029g3439 g46661g4667 g1842g3364g3038 g3404 g1842g3364g3038g3029 g3397 g3435g2025g1191g3038g3033 g3398 g2025g1191g3038g3029g3439 g46662g4667 where g1846g3364g3038 and g1842g3364g3038 are mean monthly temperature (0C) and precipitation (mm) for future periods (2016-2040); g1846g3364g3038g3029 and g1842g3364g3038g3029 are observed historic mean monthly temperature (0C) and precipitation (mm) for the baseline period (1984-2008); g2028g1191g3038g3029 and g2025g1191g3038g3029 are mean monthly temperature and precipitation coming from GCM predictions for the baseline period ; and g2028g1191g3038g3033 and g2025g1191g3038g3033 are GCM projected mean monthly temperature and precipitation for appropriate future periods. The 52 subscript k in each term represents the month. Note that future period spans from 1/1/2016 to 12/31/2040 and the baseline period spans from 1/1/1984 to 12/31/2008. Therefore, while k =1 represents Jan 1984 in the baseline period, it represents Jan 2016 in the future period. In other words, there is a one to one correspondence between the months of baseline and future periods. For example, May 2020 in the future period corresponds to may 1988 of the baseline period. Equations (1) and (2) assume that the difference in monthly averages of GCM projections between the future and baseline periods are the same as the change between observed historic monthly averages and the future monthly averages. Finally, since there are 4 GCMs and 3 green house gas emission scenarios, 12 groups of future monthly precipitation and temperature data were generated. Daily precipitation and temperature Since SWAT simulates flow and nutrients at daily time scale, monthly climate projections still need to be downscaled to daily time scale in order to study the climate change effects on hydrology and water quality. Hence, SWAT model?s stochastic weather generator, WXGEN (Sharpley and Williams, 1990) was used to generate daily rainfall from monthly statistics, such as mean monthly rain and number of wet days in that month, etc, to downscale monthly precipitation data to daily time scale. Those statistics are often estimated from historic weather records. SWAT has a built in database for such statistics compiled from long term NOAA rainfall data. For the same monthly parameters, weather generator may produce hundreds of different daily rainfall patterns, which reflects the variation of daily rainfall. In general, 20 sets is the minimum number to obtain a representative distribution of possible weather scenarios given 53 the predicted probabilities (Neitsch et al., 2005). Therefore 21 sets of daily rainfall patterns were generated by WXGEN. For a given month m, future daily precipitation was calculated as follows: g1842g3040,g3036 g3404 g1870g3040,g3036g2869 g3041 g3400 ? g1870g3040,g3036 g3041g3288 g3036g2880g2869 g3400 g1842g3364g3040 g46663g4667 Where nm is the number of days in a given month m, in this study the total number of months is 25*12=300; i is any day in the given month; g1842g3040,g3036 reflects projected daily precipitation of day i in the given month m; g1870g3040,g3036 is the generated daily rainfall for day i in the given month m by WXGEN; ? g1870g3040,g3036g3041g3288g3036g2880g2869 is the total precipitation in the given month m; g1842g3364g3040 is the future mean monthly precipitation coming from equation (1). Since there are 12 groups of g1842g3364g3040 for each month m reflecting different combinations of GCMs and green gas emission scenarios, and 21 sets of g1870g3040,g3036 generated by WXGEN, in total, 12*21=252 sets of daily precipitation data were generated. Compared with daily variation of precipitation, which may substantially affect flow and water quality, daily temperature does not fluctuate substantially in a given month. In this study, daily patterns of daily temperature were not generated; rather daily maximum and minimum temperatures for future were estimated as follows: g1846g3040,g3036g3040g3028g3051 g3404 g1846g3040,g3036g3029,g3040g3028g3051 g3397 g3435g2028g1191g3040g3033 g3398 g2028g1191g3040g3029 g3439 g46664g4667 g1846g3040,g3036g3040g3036g3041 g3404 g1846g3040,g3036g3029,g3040g3036g3041 g3397 g3435g2028g1191g3040g3033 g3398 g2028g1191g3040g3029 g3439 g46665g4667 where g1846g3040,g3036g3040g3028g3051 and g1846g3040,g3036g3040g3036g3041 are respectively daily maximum and minimum temperatures for future period (2016-2040) of a given day i in a given month m; g1846g3040,g3036g3029,g3040g3028g3051 and g1846g3040,g3036g3029,g3040g3036g3041 are daily observed maximum and minimum temperatures, respectively, for the baseline period (1984-2008); g2028g1191g3040g3029 is the mean monthly temperature from GCM predictions for the baseline period; and g2028g1191g3040g3033 is the GCM projected mean monthly temperature for the future period. Obviously, the number of daily 54 temperature patterns is consistent with the number of GCM and green house gas emission scenario combinations, which is 12 as stated before. Lastly, although each GCM under a specific green house gas emission scenario have 21 sets of daily precipitation patterns, it has only one set of daily temperature pattern as SWAT input. For instance, SWAT simulated the GFDL_cm2_0 under A1B scenario 21 times with different daily precipitation, but with the same daily temperature data. Model experiment set up: The SWAT simulations were performed for two 25-year time periods. First one was the baseline period, 1984-2008, for which calibration and validation for flow, sediment and nutrient were performed. Since there was no sufficient measured data in the Wolf Bay watershed, SWAT was calibrated and validated in the nearby data rich Magnolia River watershed. Relevant model parameters were then transferred to the Wolf Bay watershed (Wang and Kalin, 2010). This method is called regionalization approach based on spatial proximity and is widely used when there is no enough observed data in a target watershed to ensure model reliability (Merz and Bloschl, 2004; Oudin et al., 2008; Reichl et al., 2009). The second period is to simulate future climate from 2016 to 2040, for which the 3 projected LULC map and the 4 downscaled GCM under 3 green house gas emission scenarios for climate conditions were used. Parameter set was assumed to be the same as the one used for the baseline period. In order to detect the marginal and joint effects of LULC and climate change, the approach of one factor at a time was used (Li et al., 2009). This approach changes one factor at a time while holding others constant. We designed the model experiments based on that, as outlined below: 55 i) Baseline run: Most recent LULC map of 2005 and daily measured climate data (1984- 2008) from NOAA station at Robertsdale (Fig. 1) was used as SWAT input. ii) Only climate change effect: Current LULC map of 2005 and future daily climate data (2016-2040) downscaled from 4 GCM under 3 green house gas emission scenarios were used. Model had 4*3*21=252 ensemble of outputs. iii) Only LULC change effect: Three projected LULC maps for year 2030 and historic climate data (1984-2008) from NOAA station at Robertsdale were used as input. iv) Combined change effect: Three projected LULC maps for year 2030 and future daily climate data (2016-2040) downscaled from 4 GCM under 3 green house gas emission scenarios were used. Model had 252*3=756 ensemble of outputs. RESULTS AND DISSCUSSION Climate change effects on precipitation and temperature. Fig. 3 shows variations in average seasonal precipitation and temperature for the future period (2016-2040) relative to the baseline period (1984-2008). The horizontal axes in each panel indicate changes in average precipitation, while vertical axes denote changes in average temperature. The 12 dots in each panel correspond to 12 different future scenarios (4 GCMs*3 emission scenarios) with the cross indicating the average of 12 scenarios. It is clear that future climate predictions are quite different from each other, and the uncertainty range exhibit a seasonality behavior. Based on Fig. 3, all future projections indicate a rising trend in temperature, but distinct magnitudes are detected according to seasons. For summer and fall, the range is from +0.4 to +2.0 0C, while for spring and winter, the range is from +0.2 to +1.6 0C. Annual increase of mean temperature varies from +0.4 to +1.4 0C. Precipitation has a different pattern than 56 temperature. Although there is no clear trend of increase or decrease in average monthly precipitation for spring, summer and winter, in fall, 11 of the 12 scenarios show increase in precipitation, with an average of 10% for fall months. This is potentially a good thing as fall is typically the driest season in the Southeast U.S. At the annual scale, 8 of the 12 future projections predicted increase in precipitation, approximately 4% on average. Generally, based on Fig. 3, Wolf Bay watershed will more likely experience increased precipitation in future, especially in fall months. Temperature is expected to increase for all seasons, especially in summer and fall. Fig. 3 provides some general information about the potential changes in precipitation in the future in the Wolf Bay watershed, such as averages and seasonal differences. However, the change in frequency and magnitude of daily rainfall is not shown. Fig. 4 reflects exceedance probabilities for daily precipitation, which provides more insight. Probability of exceedance (PE) in the figure reflects the possibility of having rainfall amount of that magnitude or higher in a given day. Therefore, what is shown in the figure are complimentary cumulative distribution functions (CCDF). Out of the 252 complimentary CDFs generated from 252 sets of future daily precipitation data, Fig. 4 shows only the 95th and 5th percentiles (90% confidence interval). Median of the CCDF?s is also shown in the figure. It can be seen that the relative positions of projected precipitation curves and baseline curve differs with PE. Baseline fluctuates around the median curve when PE< 0.001, then falls below the 90% confidence interval from around PE=0.0015 to around PE=0.03. After PE=0.03, relative position of baseline is rising again until it becomes higher than the upper limit around PE=0.15. This means large rain events will be more intense in the future. On the contrary, the rainfall intensity of smaller events (PE<0.15) will be 57 reduced. Combining Fig. 3 with Fig. 4 one can conclude that in the future there will be a shift in rainfall pattern with large events getting more intense and smaller events becoming less intense. Climate change effects only SWAT was run at annual, monthly and daily time scales 252 times to simulate flow, sediment and nutrient loadings under future scenarios. From this ensemble of model outputs 5th and 95th percentiles were calculated to represent dry and wet conditions, respectively, in future. Median of the ensemble of model outputs was also used to represent normal conditions. Results are discussed below. Annual and monthly flow The projected annual average daily streamflow for the future period was 3.19, 4.87, and 6.81m3/s, under dry, normal and wet conditions, respectively. Compared to average daily streamflow in the baseline period, which was 4.57 m3/s, it is hard to judge if there is an increasing trend in average daily streamflow under normal conditions. We conducted a t-test for two independent samples between baseline group and future normal condition group. One-sided p-value (pooled) of 0.16 indicates no significant difference between the two groups. Therefore, the projected increase in annual average daily streamflow is statistically insignificant. For baseflow and surface runoff, the projected daily average values under the normal condition were 2.55 m3/s and 2.29 m3/s, respectively. For baseline, the corresponding annual average daily discharges were 2.33 and 2.24 m3/s, respectively. The t-test resulted in p-values of 0.07 and 0.37 for surface runoff and baseflow, respectively. Therefore, no trend was detected for baseflow under the normal condition. On the other hand, the p-value of 0.07 obtained for surface runoff is 58 not that big. Although at 5% level there is no statistical difference, at 10% there is an increase in surface runoff. Compared with Cruise?s (1999) study, our results doesn?t reflect strong decreasing trend in streamflow. Since Cruise?s conclusion was only based on the United Kingdom Hadley Center climate model and focused on the whole Southeast U.S., it is not surprising to have different predictions in streamflow. Fig. 5 shows the 5th and 95th percentiles along with the median of average monthly streamflow, surface runoff and baseflow by running SWAT with the 252 climate inputs. Under wet conditions, streamflow, direct runoff and baseflow are all showing a rising trend for all months when compared to the baseline. For dry conditions, all three reflect a declining trend, though in September and October the difference is marginal. Under the normal conditions, consistent with the rainfall, streamflow and surface runoff are projected to increase in fall months (September, October and November). Winter (December, January, and February) and spring (March, April, and May) will experience moderate increase in streamflow and surface runoff. In summer (June, July, August), no significant difference is predicted in streamflow and direct runoff. Baseflow is projected to decrease slightly in spring and summer, while increase marginally in winter. Daily flow for the whole period Fig. 6a shows the 90% confidence interval along with the median of the FDCs generated from daily streamflow by running the SWAT model with the 252 precipitation inputs. In the figure, upper and lower limit represents the wet and dry conditions, while median reflects the normal conditions in the future. According to different probability of exceedance, future daily 59 discharge illustrates substantial differences when compared to baseline. Furthermore, positions of future FDCs relative to the baseline FDC are quite similar to the precipitation CCDFs shown in Fig. 4. This is not quite surprising as precipitation is the main driver for flow. To get a better insight into the effect of climate change on daily flow responses, the relative difference of future conditions from the baseline, i.e. ([FDC of future conditions] ? [FDC of baseline situation]) / [FDC of baseline situation], were generated and depicted in Fig. 6b. Based on this figure, under wet conditions (upper confidence limit), the relative change of flow is always a positive value. It fluctuates between 75% and 10% when PE<0.01. It reaches at a plateau at 40% until PE arrives at 0.1, and then reduces to 10%. Under normal condition (median of 90% confidence interval), relative change fluctuates around 0 when PE<0.001, and reaches a maximum of about 30% around PE=0.01. It then decreases gradually from 30% to 5%, until PE=0.9. After PE=0.9, it drops drastically all the way to -30%. The shapes of these two curves are quite similar, only shifting in relative positions. Under dry conditions, as PE increases from 0 to 0.01, the relative change grows consistently from -40% to 20%. The relative change is 0 when PE is around 0.002. After PE=0.01, it decreases gradually from 20% to -15% until it reaches PE=0.85. From PE=0.85 till the end it drops drastically to -60%. When 0.002