141 12 Predictive Modeling of Lake Nitrogen, Phosphorus, and Sediment Concentrations BasedonLandUse/Land Cover Type and Pattern Pariwate Varnakovida, Narumon Wiangwang, Joseph P. Messina, and Jiaguo Qi 12.1 INTRODUCTION In watershed management and planning, one of the major problems in lakes is the need to reduce nonpoint source pollution. 1 Specic land use and land cover (LULC) types, such as “cropland” and “urban”, are associated with human activities and their physical characteristics often affect water quality. 2 Land use practices and water resources are unequivocally linked. The type and the intensity of land use have a strong inuence on the receiving water resource, especially in lakes. 3 Since agricultural land is dominant in many watersheds of the upper Midwest of the United States, it is considered a leading source for nonpoint source pollut- ants, primarily sediments and nutrients. Agricultural erosion occurs when elds are cleared of vegetation to prepare for crop planting. The physical erosion potential of some soil types, such as ne sandy loam, may be exacerbated by nonconservation agricultural practices, which may reduce the soil’s chemical fertility. The angle and length of slopes on the land also inuence the rate and amount of runoff, and in turn inuence erosion. As soil fertility declines, farmers tend to increase their application of fertilizers. This intensive application of fertilizer then becomes a major source of culturally driven eutrophication in lakes and streams. 4–6 Urbanization is another factor that impacts water quality. Catchment or subwa- tershed increases in impervious areas cause a direct impact on stream quality. 7 An impervious area is any area that no longer allows rainfall to soak into the ground, such as roads, sidewalks, rooftops, and driveways. When a site is developed, it loses its natural storage potential for rainfall. Consequently, rain that previously inltrated © 2008 by Taylor & Francis Group, LLC 142 Wetland and Water Resource Modeling and Assessment into the ground evaporates or transpires, and rain that was temporarily stored in depressions and tree canopies now rapidly runs off the site. Previous studies on lake eutrophication have focused either on external loadings or internal processes. Studies of external sources are generally limited to a single source and cannot be easily generalized. 8 Hence, to be able to provide insight into the effects of all the potential pollutant sources it is necessary to model the entire ecosystem leading to the in-lake effects. System modeling and simulation is one of the best alternatives to urban lake monitoring and water quality studies. Although several simulation models are currently being used among research- ers to predict water quality (e.g., BASINS [Better Assessment Science Integrating Point and Non-Point Sources], AGNPS [Agricultural Non-Point Source], and WASP [Water Quality Analysis Simulation Program]), most are either scale dependent or have extensive input data requirements. Therefore, for this study, we constructed the NPSSIM (nitrogen, phosphorus, and sediment simulation) model to predict TN (total nitrogen), TP (total phosporus), and TSS (total sediment) concentration in lakes based on surrounding LULC types and patterns with the express desire to manage scale and data problems, such as the interaction among local and regional processes and the lack of adequate data for model calibration. 9 Most other models simulate output based on only one continuous landscape (e.g., a watershed or a region). NPSSIM differs from other models in three main ways. First, the simulation is based on a lakeshed scale. Only the lake and the sur- rounding land that drain into it are used in the model. The model treats lakes as a single entity; therefore, lakes from different locations, such as lakes in Michigan and lakes in Wisconsin, are no different. Nutrients and sediment from different lakes across landscapes can be simulated and compared in one step. Second, the model uses landscape metrics as the main approach. Landscape metrics are known to be able to describe characteristics of land use/land cover. It is also capable of comparing LULC from different locations through the landscape neutral model. NPSSIM is therefore truly spatial scale independent because as long as DEM (digi- tal elevation model), LULC, soil type, and rainfall data are available on the same spatial resolution (e.g., 30 m for LANDSAT TM+ and 30 m NEDDEM), the model can simulate nutrient concentration regardless of the lake location. Even if spatial resolutions of the data are inconsistent, grain size adjustment is uncomplicated. Third, NPSSIM requires fewer input data than most other distributed watershed models. Although the model requires fewer inputs, it certainly incorporates hydro- logic processes by using rainfall, evaporation, and ground permeability (e.g., soil type) parameters. This model advances the state of knowledge of landscape pattern metrics as factors for the prediction of water quality parameters. Landscape indices are gen- erally widely used tools for spatial landscape analyses and serve as standards for comparison between landscapes in different parts of the world. Linking water qual- ity to landscape indices allows us to compare land and waterscapes across scales. The NPSSIM model serves to enhance our understanding of the relationships between lake water quality parameters and the wider landscape components, while providing water quality managers with a new cost-effective tool to manage water resources. © 2008 by Taylor & Francis Group, LLC Predictive Modeling 143 12.2 METHODS 12.2.1 L AND USE/LAND COVER (LULC) DATA LULC data obtained from the Michigan Geographic Data Library were derived from a classication of 2001 Landsat Thematic Mapper (TM) imagery from three seasons: spring (leaf-off), summer, and fall (senescence). 12.2.2 DIGITAL ELEVATION MODEL (DEM) A seamless 30-meter resolution DEM was produced by the United States Geologi- cal Survey (USGS). The data were checked for articial sinks that may have been caused by the production process. All sinks smaller than 8 pixels were lled. A raster grid of geographically corrected water bodies, including both streams and lakes, were overlaid on the DEM. A focal lter was used to smooth out the elevation and eliminate sinks (Figure 12.1). 12.2.3 LAKE SAMPLING METHOD The lake water quality parameter data used in this study were collected by the Mich- igan State University research group of Dr. Robert Jan Stevenson in the Department of Zoology. The data was comprised of data from 158 lakes within the Muskegon River watershed. The water quality parameter data were collected in spring and FIGURE 12.1 Digital elevation model (DEM): Michigan and the Muskegon River watershed. (See color insert after p. 162.) © 2008 by Taylor & Francis Group, LLC 144 Wetland and Water Resource Modeling and Assessment summer of 2002. The samples were ideally collected from the deepest basin of the lake. One sample was collected from each lake to represent the lake. Lake data were classied by trophic status to ensure variation in water quality from oligotrophic to hypereutrophic. Due to the lack of a completed water quality parameter dataset, 52 lakes were selected from the database to be used in model simulation. Thirty percent of the simulated dataset (17 lakes) were used for the validation. 12.2.4 LAKESHED GENERATION The lakeshed, commonly known as catchments of selected lakes, were generated using ArcINFO GRID software. To derive lakesheds, direction grids and source grids were needed. The direction grid was prepared by running the owdirection command on the DEM. The source grid is a grid that represents cells above which the contributing area was prepared using a streamlink function. Streamlink assigns unique values to sections of a raster linear network between intersections. Once the direction grid and the source grid were completed, the lakeshed was determined by using a watershed function. The lakesheds were converted to the vector data model (Figure 12.2), and boundaries were edited when necessary. Next, lakeshed boundar- ies were used to clip out the surrounding LULC and were classed to seven major categories: urban, agriculture, open land, water, golf course, forest, and wetland. 12.2.5 LANDSCAPE PATTERN METRICS In many studies, landscape pattern metrics have been used to describe changes in a landscape through time or to compare landscapes. 10 In this research, metrics were calculated on the LULC data including (1) number of patches per class per year, (2) total area per class per year, (3) mean patch size per class per year, (4) patch size standard deviation per class per year, and (5) dominance. All were calculated using the interactive data language (IDL) (TM). FIGURE 12.2 Example of lakeshed boundary: Lake Mitchell (left), Lake Cadillac (middle), and Hardy Dam pond (right). © 2008 by Taylor & Francis Group, LLC Predictive Modeling 145 12.2.6 RELATIONSHIPS AMONG LANDSCAPE METRICS AND WATER QUALITY PARAMETERS The landscape pattern metrics and slope were regressed against TP, TN, and TSS in SYSTAT 9.0 with water quality parameters (TP, TN, and TSS) being the dependent variables and the landscape parameters (landscape pattern metrics and slope) the independent variables. Metrics that produced insignicant results (P > 0.05; con- dence level 0.95) were eliminated. Finally, combinations of metrics that suggested the strongest relationships with nutrients and turbidity were used in the model. Mean patch size per year for classes including urban, agriculture, and open land were used. 12.2.7 SIMULATION MODEL The NPSSIM predictive model was developed in the STELLA 8.0 software. TN, TP, and TSS were predicted in runoff using the regression relationship between landscape and the water quality parameters (as described in section 12.2.6). Other considerable factors, such as soil permeability, rainfall, lake volume, fertilizer application rate, nutrients concentration in rainfall, nutrients outow rates, and TSS precipitate rate, were incorporated into the STELLA conceptual framework (as shown in Figure 12.3) to account for the hydrological process. Model con- stant and equations were based on regression results and Michigan Agricultural Statistics. 12.2.8 MODEL VERIFICATION AND VALIDATION After model equations were revised, and adjusting coefcients were applied where needed, it was veried by comparing simulation results with eld observations. Due to data limitations, the model was validated with TN, TP, and TSS data from 17 lakes (equivalent to 30 percent of all obtainable lake data) from the same year. These lakes were not used to create the model. Landscape metrics of the lake catchments were calculated (as described in sections 12.2.4 and 12.2.5) and entered into the model. TN, TP, and TSS were simulated in NPSSIM. The results were correlated with the in situ concentrations of the same lakes. Correlation coefcients (r 2 ) were used to report accuracy of the model. 12.2.9 SENSITIVITY ANALYSIS Sensitivity analysis was calculated on TN and TP in response to changes in mean patch size (MPS) of agriculture and urban changes. TN was more sensitive to agricul- ture when the patch size increased. Phosphorus was less sensitive to urban changes when the patch size increased. The analysis of sensitivity on fertilizer application and rainfall was performed; however, the results were not signicant. Changes in fertilizer and rainfall up to 50% of the original rate did not have a signicant impact on the water quality according to the model. © 2008 by Taylor & Francis Group, LLC 146 Wetland and Water Resource Modeling and Assessment 12.3 RESULTS A. Multivariate regression analysis suggested that the mean patch size of agri- culture and open land in the catchments had the strongest relationship with TN, implying that increasing sizes of agriculture patches release more N into lakes. The result supported the hypothesis that TN increased as agri- culture increased. #%""        $ ! " !$"# #%"" # %"" $& !"" $&  $#"  $## ! $& ! !$#!  # $#" !#""     ! $& $& # ##!# #!# ! ## # *## ,!#%! !#1 #! #! ##!# #!# .$#20  ( ) +/ !# +/ !## #! #! #!  $#" !#1 #%""'  ,!#%! ,!#%! -,!#" ! -,!$& -$&!"" -,!  - $#" !   ' #-#!# - #!# -.$#20 - $# # -#! -,! !$#! FIGURE 12.3 Conceptual nitrogen, phosphorus, and sediment simulation (NPSSIM) model. © 2008 by Taylor & Francis Group, LLC Predictive Modeling 147 B. The regression result suggested that the mean patch size of urban and open land had the strongest relationship with TP. The result did not support the hypothesis that TP would increase as agriculture increased. Urban land use seemed to have more impact on TP than agriculture. This may be because a major portion of phosphorus transport in water is by binding with other substrates. Sediment loads increase with urban land use. That, combined with the open land that creates sediment loads in runoff, could enhance phosphorus transportation into the lakes. C. TSS was most affected by the percentage of the area of urban land use and the landscape slope. The regression result supported the hypothe- sis that TSS increased as urban area increased. Urban built-up material created impervious surfaces, which increased velocities in surface run- off. Higher slopes also enhanced the process. Clearly increasing speeds in surface runoff increases water carrying capacity for particles and sediment. D. Validation of the model between observed and predicted TP, TN, and TSS resulted in r 2 of 0.4535, 0.2471, and 0.4593, respectively (Figure 12.4). 12.4 CONCLUSION AND DISCUSSION Lake ecosystems are very complicated and direct relationships among TN, TP, TSS, and water quality parameters are extremely difcult to determine. Correlation coef- cients from our multivariable regression models were low (r 2 < 0.25); however, with some other variables, such as slope, soil permeability, rainfall, nutrients con- centration in rainfall, and fertilizer included into the model, the prediction accuracy improved (r 2 0.25–0.46). Nonlinear or stepwise regression may capture the changes and improve the predictability. A stochastic model, as opposed to our deterministic NPSSIM model, may be an alternative approach because many variables (e.g., rain- fall, fertilizer application) are not constant in nature. Even though the validation correlation coefcient did not show impressive accuracy, the nutrient models did capture the trend of how water quality param- eters respond to changes in LULC (bar graphs in Figure 12.4). The TN increased as agricultural areas increased as a result of the intensive application of fertilizer use. The TP increases as a catchment or subwatershed increases in impervious area with a direct impact on stream quality increases. The TSS model seemed to predict the average in all lakes. However, the percentage of urban area was selected as the signicant variable in the regression process, which shows that the urban area was more correlated to TSS than other LULC types. Velocities in surface runoff were increased by impervious surfaces in urban land. On the other hand, area that had been classied as agricultural land was mostly covered with vegetation; therefore, a smaller amount of sediment was washed out of the surface. Additional param- eters may need to be included in the model to enhance the response of TSS to dif- ferences in LULC. Sensitivity analysis shows an insignicant relationship between water quality parameters and fertilizer application. This was unexpected. The model may lack variables that better link fertilizer to the entire system. Further revision is needed. © 2008 by Taylor & Francis Group, LLC 148 Wetland and Water Resource Modeling and Assessment Without a time constraint, the model could potentially be improved. Rened equations and constants for the model could be developed. The input parameters were based on available data and may have reduced model predictability. This proj- ect and model serve as preliminary work for future research. TP R 2 = 0.4535 0 10 20 30 40 0 102030405060 0 10 20 30 40 Lakes TP TP TN R 2 = 0.2471 0 200 400 600 800 1000 L_HESS_0403 L_FREM_0502 L_GRAS_0602 L_TAMA_0403 L_BEAR_0502 L_LILY_0602 L_RYER_0602 L_KIMB_0502 L_MITC_0602 L_TOWN_0502 L_PLEA_0502 L_CLME_0602 L_CROO_0602 L_WELL_0602 L_HAYM_0602 L_SILV_0602 Predicted Predicted 0 300 600 900 1200 Lakes TN TN Predicted Observed Predicted Observed TSS R 2 = 0.4593 0 2 4 6 8 02468101214 200 400 600 800 1000 1200 Observed Predicted Observed Predicted Observed Predicted Predicted 0 3 6 9 12 15 Lakes TSS TSS L_TODD_0502 L_BEAR_0502 L_HESS_0403 L_TAHA_0403 L_GRAS_0602 L_KIHP_0502 L_WELL_0602 L_TOWN_0502 L_PREH_0502 L_TODD_0502 L_HITC_0602 L_HAYH_0602 L_RYER_0602 L_PLEH_0502 L_CLHE_0602 L_CROO_0602 L_SILY_0602 L_LILY_0602 L_BEAR_0502 L_HESS_0403 L_FREM_0502 L_GRAS_0602 L_CROO_0602 L_TAMA_0403 L_PLEA_0502 L_KIMB_0502 L_TOWN_0502 L_RYER_0602 L_LILY_0602 L_TODD_0502 L_HAYM_0602 L_SILY_0602 L_MITO_0602 L_OLME_0602 L_WELL_0602 Predicted Observed FIGURE 12.4 Validation results. © 2008 by Taylor & Francis Group, LLC Predictive Modeling 149 REFERENCES 1. Copeland, C. 1999. Clean water action plan: Background and early implementation. Washington, DC: CRS Report, 98–150. 2. Bianchi, M., and T. Harter. 2002. Nonpoint sources of pollution in irrigated agricul- ture. Farm water quality planning reference sheet 9.1. Oakland: University of Califor- nia, 1–8. 3. Institute of Water Research. Michigan State University. http://www.iwr.msu.edu/ edmodule/water/luwhome.htm. 4. Carpenter, S. R., N. F. Caraco, D. L. Correll, R. W. Howarth, A. N. Sharpley, and V. H. Smith. 1998. Nonpoint source pollution of surface waters with phosphorus and nitro- gen. Ecological Applications 8(3):559–568. 5. Tinker, P. B. 1997. The environmental implications of intensied land use in develop - ing countries. Philosophical Transactions: Biological Sciences 352(1356):1023–1032. 6. Mueller, D. K., and D. R. Helsel. 1996. Nutrients in the nation’s waters—too much of a good thing? U.S. Geological Survey Circular 1136. Washington, DC: U.S. Geological Survey. 7. Roa-Espinosa, A., J. M. Norman, T. B. Wilson, and K. Johnson. 2003. Predicting the impact of urban development on stream temperature using a Thermal Urban Runoff Model (TURM). In Proceedings of the National Conference on Urban Stormwater: Enhancing Programs at the Local Level, Chicago, Illinois, February 17–20, 2003. Washington, DC: U.S. Environmental Protection Agency, 369–389. 8. Puijenbroek, P. J. T. M., J. H. Janse, and J. M. Knoop. 2004. Integrated modeling for nutrient loading and ecology of lakes in the Netherlands. Ecological Modelling 174:127–141. 9. Shanahan, P., M. Henze, L. Koncsos, W. Rauch, P. Reichert, L. Somlyody, and P. Vanrolleghem. 1998. River water quality modelling: II. Problems of the Art. Water Science and Technology 38(11):245–-252. 10. Iverson, L. R. 1988. Land-use changes in Illinois, USA: The inuence of landscape attributes on current and historic land use. Landscape Ecology 2:45–62. © 2008 by Taylor & Francis Group, LLC . such as “cropland” and “urban”, are associated with human activities and their physical characteristics often affect water quality. 2 Land use practices and water resources are unequivocally. prediction of water quality parameters. Landscape indices are gen- erally widely used tools for spatial landscape analyses and serve as standards for comparison between landscapes in different parts. RELATIONSHIPS AMONG LANDSCAPE METRICS AND WATER QUALITY PARAMETERS The landscape pattern metrics and slope were regressed against TP, TN, and TSS in SYSTAT 9.0 with water quality parameters (TP, TN, and