## 1. Introduction

Water pollution has received increasing attention, and many countries have increased their investments into water pollution control and water resources protection [

1]. Designing scientific, reasonable, and efficient management measures to control or reduce pollutants at the watershed scale has become one of the most challenging problems for policy researchers and decision makers [

2]. Many policies for the selection of best management practice (BMPs) have been created and applied to specific cases all around the world. For example, the United States and Europe have developed the corresponding Total Maximum Daily Loads and European Water Framework Directive for such purposes [

3,

4]. The implementation of these plans has provided a sufficient theoretical basis for subsequent watershed governance research [

5].

BMPs are the most effective measure for controlling watershed pollution, including vegetative filter strips, land-use transformation, reducing the amount of fertilizer, terraces, and so on [

6]. In general, BMP implementation plans should consist of a combination of maximum pollution reduction and minimal financial costs, due to limited budgets [

7]. To our best knowledge, there are three optimization techniques which achieve the purpose. The first is setting a fixed number of scenarios manually and then calculating the corresponding pollutant production and cost separately [

8,

9]. By comparing the results of a limited number of different BMP scenarios, the final solution can be picked out. This method is straightforward and easy to implement, but may cause biased results as it depends on the experience of managers. Thus, the solution may not be the most cost-effective at the watershed scale [

10]. The second is aggregating the environmental goals and economic factors into a single compromised objective function (e.g., a genetic algorithm or TaBu search algorithm) [

11,

12]. Through coupling the watershed model and optimization algorithm, only one optimal solution can be searched [

13]. Compared with the first method, this method is more objective but usually takes more time, due to the necessary model runtime for each population per generation. The last technique is the coupling of a multi-objective optimization algorithm and a distributed watershed model to search over a set of solutions. This technique is similar to the second one, but it is able to provide a range of different trade-off BMPs among two or more conflicting objective functions. Due to its comprehensiveness and accuracy, it has been widely used in recent years [

14].

The non-dominated sorting genetic algorithms II (NSGAII) is the most popular method for multi-objective optimization, whose ultimate goal is to find “Pareto-optimal” solutions, which is a modified version based on the genetic algorithm [

15]. For example: Maringanti et al. utilized the Soil and Water Assessment Tool (SWAT) model and NGSAII to analyze the funding input under different combinations of fertilization reduction ratios, riparian filter belt widths, and other agricultural management measures in a tributary of the Mississippi River [

16,

17]; Ahmadi et al. combined the SWAT model and the NSGAII algorithm to evaluate the prevention and treatment effects of Atrazine through various evaluation indicators and non-point source management measures in the Eagle Creek Watershed in Indiana, USA [

18]; and Geng et al. coupled the SWAT model and NSGAII algorithm to calculate the relationship between the amount of nutrient reduction and the required funding in the Chaohe River Watershed upstream of the Miyun Reservoir in China [

19]. However, almost all studies in this category have focused on non-point source pollution BMPs and ignored point source BMP measures. It can be inferred, therefore, that the description of point source measures is not easy in the management practices of the watershed model.

Besides the selection of optimization technique, effective watershed management requires an understanding of the fundamental hydrologic and physicochemical processes in the watershed system, which are non-linear, dynamical, and complex [

20]. Therefore, the applicability of the basin watershed model is very essential. A number of comprehensive watershed models have been developed to simulate hydrology and water quality in basins, and previous studies have demonstrated that some watershed models are well-behaved for the selection and targeted placement of BMPs (e.g., SWAT and AnnAGNPS) [

21,

22]. However, the watershed models in most previous studies have a high demand for data, and are difficult to apply in some areas where there is a lack of data [

23]. The Revised Generalized Watershed Loading Function (RGWLF) is an improved semi-distributed hydrological model based on the Generalized Watershed Loading Function (GWLF) model, which has favorable stability, robustness, and less data requirements [

24].

Given the above considerations, in this study, we incorporated RGWLF and NSGAII to identify a set of optimal BMPs based on both point and non-point source pollution control practices. Three tasks were completed to accomplish this research target: (1) adding the nutrient channel routing algorithms into the RGWLF then calibrating and verifying the parameters of the model; (2) determining the specific point source and non-point source management measures based on the established model; and (3) coupling the RGWLF model and NSGAII optimization algorithm based on a parameter sensitivity analysis to identify the optimal spatial allocation of BMPs for dissolved nitrogen.

## 4. Conclusions

In this study, we supplemented the RGWLF model with the addition of a pollutant attenuation process in the river channel and a bank filter strips reduction function, based on our previous research. Meanwhile, taking into account the regional pollution source composition of the watershed studied and the characteristics of the models used, three types of pollution control measures—point source wastewater treatment, bank filter strips, and converting farmland to forest—were considered, and the cost of each measure was determined. Furthermore, the optimization algorithm NGSAII was linked with the RGWFL watershed model and the implemented measures to search for a Pareto-optimal set of BMPs. Before the final optimization calculation, sensitivity analyses for the four key parameters of NGSAII were performed. According to the results of the sensitivity analysis, the entire coupled model system supplied 50 solutions which could provide managers with many pollution reduction options. In the end, depending on the cost, we chose four gradient solutions and demonstrated the geospatial distribution of their different management measure characteristics and briefly analyzed the differences and features of the four solutions.

The research results show that, with almost no increase in data requirements, dissolved nitrogen had an excellent simulation performance, expanding the spatial scope of pollutant simulation for the GWLF. Moreover, due to its robustness and semi-distribution, the RGWLF model was able to be coupled with NSGAII. The entire linkage system had good performance in the optimization process and provided a range of watershed implementation measures for DisN reduction and minimizing cost, which is a worthy reference for policy researchers and decision-makers to realize their watershed management goals. However, limited by the available data, funding, and researchers’ capabilities, we did not consider other target indicators, such as sediment, phosphorus, and so on, which could be carried out in future research.