This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).

Suites of Best Management Practices (BMPs) are usually selected to be economically and environmentally efficient in reducing nonpoint source (NPS) pollutants from agricultural areas in a watershed. The objective of this research was to compare the selection and placement of BMPs in a pasture-dominated watershed using multiobjective optimization and targeting methods. Two objective functions were used in the optimization process, which minimize pollutant losses and the BMP placement areas. The optimization tool was an integration of a multi-objective genetic algorithm (GA) and a watershed model (Soil and Water Assessment Tool—SWAT). For the targeting method, an optimum BMP option was implemented in critical areas in the watershed that contribute the greatest pollutant losses. A total of 171 BMP combinations, which consist of grazing management, vegetated filter strips (VFS), and poultry litter applications were considered. The results showed that the optimization is less effective when vegetated filter strips (VFS) are not considered, and it requires much longer computation times than the targeting method to search for optimum BMPs. Although the targeting method is effective in selecting and placing an optimum BMP, larger areas are needed for BMP implementation to achieve the same pollutant reductions as the optimization method.

Nonpoint source (NPS) pollution from agricultural watersheds has become one of the major water quality concerns [

The adverse impacts from agricultural areas can be controlled by implementing best management practices (BMPs) to reduce source or retard pollutant transports in a watershed. Many studies have used simulation models to evaluate BMP effectiveness and determine the optimum BMPs to improve water quality at the farm level [

Targeting is a plan-based method to place BMPs in critical source areas which contribute a disproportionate amount of NPS pollutants. Many studies have been conducted to identify critical source areas and to estimate the improvement of water quality due to implementation of selected BMPs in those critical regions [

Optimization studies related to selection of BMPs have traditionally used cost minimization as an objective function. Cost as an optimization function does not ensure that the watershed areas under BMPs are also minimized. Many researchers have indicated that a relatively small portion of a watershed contributes a larger amount of pollutants. BMPs are generally targeted in those high-risk watershed areas. To enhance the effectiveness of BMPs, achieving the same pollutant reductions with less areas should be considered. In order to compare the watershed areas that needs to have BMPs under targeting and optimization options to produce similar water quality benefits, we used the watershed area as one of the objective functions in this study. Therefore, the overall goal of this study was to compare the selection and placement of optimal BMPs using an optimization model with various BMP options and a targeting method for achieving a high level pollutant reduction with BMP implementation in a small portion of the pasture lands. The hypotheses we tested were as follows (1) selection and placement of BMPs from different sets of BMP options using a genetic algorithm (GA) optimization tool can result in different water quality improvements; (2) Limiting the BMP options to the BMPs which have a greater pollutant reduction rate can assist the optimization tool to allocate BMPs more effectively. We used the multiobjective optimization model developed by Maringanti

This study was conducted in the Lincoln Lake CEAP watershed, a 32 km^{2} agricultural watershed within the Illinois River basin located in Northwest Arkansas and Eastern Oklahoma (

Moores Creek and Beatty Branch are the two major tributaries in the Lincoln Lake watershed representing 21 and 11 km^{2} of the watershed area, respectively. The watershed has a mixed land use with pasture, forest, urban residential, urban commercial and water representing 35.8%, 48.6%, 11.9%, 1.5% and 2.2% of the watershed area, respectively (

Location of Beatty Branch, Moores Creek, land-use distribution and the gauging stations in the Lincoln Lake watershed.

The Soil and Water Assessment Tool (SWAT, version SWAT 2009), was used to estimate the effectiveness of various BMP combinations in reducing pollutant losses in a previous study [

The SWAT model has the ability to define specific types of manure and fertilizers by building fertilizer and manure components, such as fractions of mineral N (P), organic N (P), and a ratio of ammonium nitrate to mineral N in the SWAT fertilizer database. The pasture management information, including amount of litter and fertilizer application, timing of manure and fertilizer application, grazing intensity and dates were obtained from a detailed review of historical nutrient management plans and interviews with 63 out of 75 farmers in the watershed [

Model calibration and validation were performed for monthly stream flow, total sediment (TS), total nitrogen (TN) and total phosphorus (TP) using the measured flow and water quality data collected at the Upper Moores Creek for the period January 1996–February 1999, January 2000–December 2003 and January 2006–December 2007. A total of 10 SWAT parameters were calibrated using Nash-Sutcliffe efficiency (NSE) [^{2}) as the model performance criteria. Detailed information of calibrated SWAT parameters and performance of the SWAT model can be found in Chiang

The watershed BMPs considered in this study were grouped into three categories: grazing and pasture management, vegetated filter strips, and nutrient management. These scenarios were based on detailed interactions with the watershed stakeholders and history of past BMPs implemented in the watershed [

Vegetated filter strips (VFS) have been proven to be an effective management practice for trapping sediment and nutrients in field runoff [

Nutrient management scenarios evaluated in this study included poultry litter application rates, litter characteristics, and application timing. The litter application rates evaluated were 1, 1.5 and 2 tons/acre in spring (applied on 30 April) and summer (31 August) to support growth of warm season grasses, and 2, 2.5 and 3 tons/acre in fall (15 October) to support growth of cool season grasses. For all application rates and timings evaluated in this study, two types of poultry litter were selected—normal poultry litter and alum-amended litter.

A total of 171 BMP combinations were simulated using the SWAT2009 model with dynamic land use changes during 1990–2007 in a previous study [

A genetic algorithm (GA) is a search technique to find solutions for optimization problems. Genetic algorithms are based on techniques inspired by evolutionary biology such as inheritance, selection, crossover and mutation. An algorithm is started with a set of solutions (chromosomes), called a population. The initial population of chromosomes is randomly generated for the given population size (

Overview of the GA method.

Multiobjective optimization problems have been evaluated in the hydrology/water quality field, where optimal decisions need to be taken between two or more conflicting objectives. Single-objective optimization yields a single optimal solution, while a multiobjective optimization produces a family of near-optimal solutions known as Pareto-optimal set. Deb _{pol, hru} is the unit pollutant load from a _{pol, bmp} is the pollutant reduction efficiency of BMP,

A BMP tool was used to provide pollutant effectiveness for each BMP that can be implemented at a

Four parameters for a GA optimization are population size, number of generations, crossover rate and mutation probability. Population size determines the number of solutions considered for the evolutionary process. Crossover rate and mutation probability are critical in the optimization process in terms of creating a new set of child population which might be stronger than the parent population and eliminating the weaker individuals. The optimization process continues until a given number of iterations known as generations. Generally, the larger the population size, the more spread the solution space. Increasing the number of generations can also improve the performance of GA. However, it also increases the computing time to reach the near-optimal solution.

A sensitivity analysis of four GA parameters for different allele sets (BMP options) was performed to determine the influence of the parameters on the Pareto-optimal front and to identify the optimal parameter values. In order to evaluate the individual influence of a GA parameter on the Pareto-optimal front, one parameter (population size, number of generation, crossover and mutation probability) was changed at a time and the other parameters remained as default values (

Default and optimal GA parameters for different allele sets selected from sensitivity analysis.

Parameter | Population | Number of Generations | Crossover Probability | Mutation Probability |
---|---|---|---|---|

Default | 100 | 1,000 | 0.9 | 0.0001 |

Optimal for different allele sets | ||||

171 BMPs (All) | 5,000 | 40,000 | 0.9 | 0.001 |

NG BMPs | 3,000 | 40,000 | 0.5 | 0.001 |

OG BMPs | 5,000 | 40,000 | 0.9 | 0.001 |

SP BMPs | 3,000 | 40,000 | 0.7 | 0.001 |

SU BMPs | 3,000 | 40,000 | 0.9 | 0.001 |

VFS0 BMPs | 5,000 | 40,000 | 0.7 | 0.001 |

VFS42 BMPs | 3,000 | 40,000 | 0.5 | 0.001 |

The targeting method was chosen as a comparison of the selection and placement of BMPs from the GA optimization tool. The pasture HRUs were first ranked by the TN or TP losses. A single BMP scenario that has the greatest TN or TP reduction rate is implemented on the top ranked HRUs which accounted for 20%, 40%, 60%, 80% and 100% of the total pasture area. The pollutant losses from the entire pasture lands were calculated by summing up the pollutant losses from the HRUs that have the selected BMP scenario implemented and the current pollutant losses from the rest of pasture HRUs, and then divided by the total pasture area. One of the differences between GA optimization method and targeting method is that GA requires much longer computing time for a larger population size (>5,000) to get a wide range in solution space (

The optimal GA parameters were selected using the sensitivity analysis of the optimal front. A total of seven sets of sensitivity analyses were performed for 171 BMPs. The sensitivity of GA parameters for the optimization model with a set of 171 BMP options (All), namely, population size, number of generations, crossover probability, and mutation probability, are shown in

Pareto-optimal fronts for the sensitivity analysis of genetic algorithm (GA) parameters for the optimization model with a set of 171 BMP options (All).

As the population size increased from 10 to 5,000, more individuals were present during each evolution and there was a higher probability of obtaining a better offspring. Therefore, an improvement of the Pareto-optimal front that is getting closer to the origin was observed during the increases of population size. For the 171 BMP options, when population size further increased to 5,000 the individuals in the solution space had more freedom in terms of more spread of solutions compared to other population sizes (

Pareto-optimal fronts for the sensitivity analysis of genetic algorithm (GA) parameters for the optimization model with a set of VFS42 BMP options.

Similar to the population size, an increase in the number of generations can lead the Pareto-optimal front closer to the origin and a better optimal solution can be found. The larger the number of generations is, the better the fittest individuals for reproduction can be selected. The Pareto-optimal front greatly improved when the number of generations increased from 100 to 1,000, while there was no considerable change between 10,000 and 40,000 generations (

Unlike the population size and the number of generations, an increase in crossover probability did not always result in a better Pareto-optimal front. For example, the solutions of the model with the 171 BMP options improved when the crossover probability increased from 0.1 to 0.4, but the front moved away from the origin when it further increased to 0.5 and 0.7 (

No consistent pattern in the shift of the Pareto-optimal front was found for the mutation probability. For both the models with sets of 171 BMP and VFS42 BMP options, a slightly higher mutation probability (0.001) than the default value (0.0001) made the Pareto-optimal front move toward the origin (

An interesting result was observed when comparing the Pareto-optimal fronts for these models with different BMP options using their optimal GA parameters (

Comparison of the Pareto-optimal fronts for models with different BMP options using their optimal GA parameters.

After assessing the sensitivity analysis for the GA parameters, various final values of the GA parameters were applied to each model to search for the optimal BMP solutions (

A total of 461 pasture HRUs were ranked by the pollutant losses. The BMPs that resulted in the greatest TN reduction rate are the BMP combination of buffer strips (

Solutions which are medians of the range of pollutant loads and BMP-implemented area for different optimization models.

BMP Options | TN Load (kg/ha) | BMP Area(%) | TP Load (kg/ha) | BMP Area(%) |
---|---|---|---|---|

ALL | 3.39 | 0.77 | 0.74 | 0.77 |

NG | 3.36 | 0.82 | 0.70 | 0.82 |

OG | 3.41 | 0.75 | 0.74 | 0.75 |

SP | 3.51 | 0.78 | 0.75 | 0.78 |

SU | 3.77 | 0.74 | 0.79 | 0.74 |

VFS0 | 3.75 | 0.82 | 1.11 | 0.82 |

VFS42 | 3.38 | 0.77 | 0.77 | 0.77 |

Note: ALL denotes all 171 BMP options; NG denotes the BMP options containing only BMPs with no grazing management; OG denotes the BMP options containing only BMPs with optimum grazing management; SP denotes the BMP options containing only BMPs with spring litter application; SU denotes the BMP options containing only BMPs with summer litter application; VFS0 denotes the BMP options containing only BMPs with no buffer strips; VFS42 denotes the BMP options containing only BMPs with buffer strips with a VFS ratio of 42.

Comparison of the Pareto-optimal fronts of different optimization models after the final generation and the solutions obtained from the targeting method for total nitrogen (TN) and total phosphorus (TP) reduction.

The optimal solution for the model with 171 BMP options was distributed throughout the watershed (

Compared to the distribution of the BMP-implemented area from the optimization tool, a slightly different distribution of the selected BMP in the watershed by using the targeting tool was observed (

Distribution of BMPs selected by the optimization model with the 171 BMP options on 77% of the pasture lands in the watershed.

Distribution of BMPs selected by the targeting method on 77% of the pasture lands in the watershed.

In order to compare the performance of optimization models with different BMP options, various solutions that meet either the optimal pollutant reduction or BMP-implemented area by the model with 171 BMP options were selected (

When only BMP combinations that consist of litter application in spring were selected for the optimization model, more pasture lands (89%) are needed to achieve the same pollutant reduction from the optimization model with the 171 BMP options (

Solutions from other models which at least meet the same pollutant reduction of the model with 171 BMP options.

BMP options | TN Load (kg/ha) | BMP Area(%) | TP Load (kg/ha) | BMP Area(%) |
---|---|---|---|---|

Baseline | 4.55 | 0.00 | 1.66 | 0.00 |

ALL | 3.39 | 0.77 | 0.74 | 0.77 |

NG | 3.39 | 0.80 | 0.72 | 0.80 |

OG | 3.39 | 0.77 | 0.73 | 0.77 |

SP | 3.39 | 0.89 | 0.68 | 0.89 |

SU | 3.54 | 1.00 | 0.66 | 1.00 |

VFS0 | 3.62 | 1.00 | 1.07 | 1.00 |

VFS42 | 3.39 | 0.75 | 0.78 | 0.75 |

Many studies have been conducted to optimize the selection and placement of BMPs to economically reduce pollutant loads from watersheds by using plan- or performance-based methods. The objectives of this study were to: (1) compare the selection and placement of BMPs using a genetic algorithm (GA) optimization and a targeting method; (2) evaluate the impacts of various BMP options

Distribution of selected BMPs from different optimization models to meet the same pollutant reduction of the model with the 171 BMP options.

on the optimal solutions from optimization. Two objective functions were used to minimize the TN and TP losses and the BMP-implemented pasture area. It was found that optimization required much longer computation time than the targeting method to obtain a more spread of solutions. The solutions obtained from the optimization tool were optimal for both reducing TN and TP losses by placing BMPs in the same pasture areas, while the targeting method focused on reducing one individual pollutant loading at a time by placing a single suite of BMPs in different areas, which may not be practical due to various land characteristics or farmers’ choices of BMPs. Overall, when using the targeting method more pasture areas are needed to have BMPs implemented in order to achieve the same pollutant reductions that result from the optimal BMPs selected by optimization.

A total of 171 BMP scenarios were grouped by no grazing (NG), optimum grazing (OG), spring litter application (SP), summer litter application (SU), no buffer strips (VFS0) and buffer strips with a VFS ratio of 42 (VFS42) as various sets of BMP options for evaluating their impacts on the optimal solutions from the optimization model. The results showed that limiting the BMP options to certain BMPs, such as buffer strips with a VFS ratio of 42, could result in greater pollutant reductions within smaller pasture areas managed with BMPs. However, when only summer litter application or no buffer strips are considered during optimization and the optimal BMPs are implemented in the entire pasture areas, they still resulted in greater pollutant losses than the solutions from the model with 171 BMP options. Therefore, it is essential to carefully select the BMP options for optimization in order to obtain more effective solutions in minimizing pollutant losses and BMP-implemented area in a watershed. Moreover, for a more comprehensive evaluation of selection and placement of BMPs in a watershed, other pollutants of concerns, and cost and maintenance of selected BMPs options should be taken into consideration when applying this evaluation framework.

Li-Chi Chiang and Indrajeet Chaubey designed research; Li-Chi Chiang conducted research, analyzed data and drafted this manuscript under Indrajeet Chaubey’s supervision; Chetan Maringanti developed the GA model for the research and assisted the analyses; Tao Huang managed the manuscript and provided some useful comments; Li-Chi Chiang had the primary responsibility for the final content. All authors have read and approved the final manuscript.

The authors declare no conflict of interest.