Forecasting and Fertilization Control of Agricultural Non-Point Source Pollution with Short-Term Meteorological Data

Abstract

Agricultural non-point source pollution (AGNPSP) is one of the core challenges facing global water environment management. Existing research mainly focuses on post-event estimation of pollution loads and source analysis, while studies on proactive risk warning for watershed non-point source pollution are relatively limited, especially those that integrate with agricultural production practices. Therefore, this study takes the River Tongyang Watershed as the research object and establishes a fertilization warning and regulation model based on short-term meteorological data. First, it simulates the migration and transformation processes of pollutants within the watershed under different meteorological conditions and analyzes their spatiotemporal evolution characteristics. Then, combined with real-time water quality monitoring data at the lake inlet, it calculates the residual environmental capacity for pollutants in the river water. Finally, based on this environmental capacity and the farmland area, it back-calculates the maximum safe fertilization amount for each plot under different meteorological scenarios to achieve precise fertilization management. When the planned fertilization amount does not exceed this maximum safe value, environmental risks are within a controllable range; if exceeded, fertilization should be proportionally reduced to prevent non-point source pollution. The results indicate that this model can accurately predict the concentration trends of non-point source pollutants and can develop differentiated fertilization strategies based on rainfall scenarios. The “fertilization determined by water” decision-making framework established in this study provides a technically significant pathway for shifting watershed agricultural non-point source pollution management from passive treatment to active prevention.

1. Introduction

Agricultural non-point source pollution (AGNPSP) is mainly caused by the leakage of organic or inorganic nutrients such as nitrogen and phosphorus fertilizers and pesticides from surface runoff of farmland or agricultural wastewater discharge [1]. Its sources are diverse, including domestic sewage, fertilizer application, and household waste [2]. Compared with point source pollution, non-point source pollution is characterized by dispersed distribution, complex tracing, and significant differences in characteristics among different land types during flood and non-flood periods, which brings great challenges to its monitoring and management [3]. In China, nitrogen and phosphorus from agriculture account for 81% and 93% of total water pollution, respectively, making them the main contributors to water eutrophication [4].

In the field of research on agricultural non-point source pollution, scholars at home and abroad have carried out extensive explorations. Internationally, early studies mainly focused on risk assessment of pollutant loads and the analysis of pollution sources. The “3S” technologies (GIS, RS, GPS) were widely used to explore the relationship between AGNPSP and soil erosion, fully demonstrating the advantages of “3S” technologies in non-point source pollution assessment [5,6,7,8]. Hussain et al. [9], based on an assessment of existing major non-point source pollution models, emphasized the necessity of establishing comprehensive, real-time monitoring systems to more effectively address AGNPSP issues and promote sustainable agricultural development. In recent years, building on international models, domestic scholars have also conducted extensive research tailored to China’s complex geographical and agricultural cultivation conditions. Zhang et al. [10] discovered through studying the nitrogen and phosphorus output of the Lanlingxi small watershed during the rainy season that rainfall during periods of abundant water is the main driving force for nitrogen and phosphorus pollution output in the soil. During this time, the total nitrogen and total phosphorus losses carried away by runoff accounted for 88% and 90% of the total nitrogen and phosphorus loss in the rainy season, respectively, indicating a significant positive correlation between rainfall intensity and nitrogen-phosphorus loss. Gao et al. [11], through analyzing the sources of nitrogen and phosphorus pollution in typical rural watersheds of Chaohu, found that the output of non-point source pollution such as nitrogen and phosphorus is not only influenced by terrain slope and land use practices but also regulated by factors such as rainfall. Wang et al. [12] studied the nitrogen and phosphorus discharge characteristics of the Tuojiang agricultural small watershed in 2012–2013 and showed that from July to September 2012, due to concentrated and heavy rainfall, nitrogen and phosphorus losses during this period were significantly higher than in the same period of 2013. Rainfall not only directly washed away nitrogen and phosphorus nutrients in the soil through surface runoff but also may enhance the mobility of dissolved nutrients in the soil during periods of abundant hydrology, thereby exacerbating nutrient loss loads. Therefore, it can be seen that rainfall is a key environmental factor driving agricultural non-point source pollution migration, and its spatiotemporal distribution characteristics and the impact of extreme rainfall events must be fully considered in pollution load estimation and control strategies.

As one of China’s five major freshwater lakes, Chaohu plays a key role in regional development, serving multiple functions such as water supply, flood control, irrigation, fisheries, and tourism [13]. However, with the expansion of agricultural production, the use of fertilizers and pesticides has increased significantly. These chemicals enter rivers through surface runoff during precipitation and irrigation, leading to water pollution [14]. The northern and southern regions of Chaohu are widely cultivated, with extensive agricultural areas and dense river networks, making non-point source pollution particularly prominent [15]. Among them, the River Tongyang Watershed, as a typical representative of the Chaohu Watershed, has long been severely affected by agricultural non-point source pollution. This watershed faces issues such as weak water resource management, inadequate supervision of agricultural discharge, and low environmental awareness among farmers, which further exacerbate the spread of pollution. Nevertheless, existing research has largely focused on post-event assessments of rainfall impacts on pollution output, lacking effective methods for dynamically incorporating weather forecast information into agricultural non-point source pollution risk warnings. Therefore, this study attempts to construct a pollution warning and fertilizer management model that integrates short-term rainfall forecasts, aiming to provide a scientific basis for precise prevention and control of agricultural non-point source pollution and to establish a scientific pollution monitoring and warning system.

2. Materials and Methods

2.1. Research Area

The River Tongyang Watershed is located on the northern shore of Lake Chaohu and is a small sub-watershed within the Lake Chaohu watershed (Figure 1). It covers several surrounding villages and agricultural areas, with a total watershed area of approximately 90 km². The topography of the watershed is characterized by high terrain in the northwest and low terrain in the southeast. The upstream area is mainly composed of hilly ridges with relatively undulating terrain; the mid- and downstream areas are densely populated town regions with intensive land use; the downstream section of the River Tongyang enters the plains embankment area, showing typical lowland plain features. This terrain and landform are representative within the Lake Chaohu Watershed, affecting water distribution, agricultural production patterns, and the ecological environment in the region. An automatic water quality monitoring station is set up near the middle and lower reaches of the River Tongyang (117°38′10.26″ E, 31°39′40.20″ N) to obtain water quality data of the river section entering the lake. The monitoring station is located approximately 700 m from the mouth of the River Tongyang entering the lake, effectively reflecting the water quality condition before the river flows into the lake. The River Tongyang Watershed has a subtropical humid monsoon climate, with an average annual temperature of about 16 °C and annual precipitation around 1030 mm, indicating sufficient rainfall. Land use in the watershed is mainly forest and farmland, with the primary crops being rice, wheat, and corn. With the expansion of agricultural scale, the reliance on fertilizers and pesticides in the watershed has increased year by year. The increased use of fertilizers has led to some pollutants entering water bodies through surface runoff, especially during the flood season when rainfall causes a large amount of pollutants to flow into the River Tongyang, resulting in severe eutrophication. Therefore, studying the characteristics and composition of non-point source nitrogen and phosphorus pollution in the River Tongyang Watershed, monitoring nitrogen and phosphorus concentration changes, and guiding agricultural fertilization and pollution control are of significant practical importance for improving water quality, managing rural non-point source pollution, and preventing eutrophication in Lake Chaohu.

2.2. Introduction to Data Sources

This study employs multi-source data to systematically analyze the issue of agricultural non-point source pollution in the River Tongyang Watershed. The data used primarily include digital elevation model [16] (DEM, Copernicus GLO-30) data (Figure 2a), land use types in 2021 [17] (Figure 2b), watershed rainfall, river attributes, meteorological and hydrological information, cultivated land area, and average terrain slope. Land use and DEM data were obtained from the European Space Agency (ESA). Information on the cultivated land area, hydrological data, river attributes, historical rainfall, and average slope within the watershed was provided by the Nanjing Institute of Geography and Limnology, Chinese Academy of Sciences. Meteorological data were obtained from the website of the Central Meteorological Observatory, while annual and monthly rainfall data were cited from the Chaohu Statistical Yearbook [18]. Water quality data at the River Tongyang inflow to the lake were sourced from an automatic water quality monitoring station located in the middle and lower reaches of the river. This station was constructed and operates strictly in accordance with the “Technical Specifications for Automatic Monitoring of Surface Water (Trial)” (HJ915—2017), with a monitoring frequency of once every four hours and indicators including total nitrogen, ammonia nitrogen, total phosphorus, chemical oxygen demand, pH, and water temperature. Field investigations indicated that there are no large industrial enterprises within the River Tongyang Watershed; the region contains only one wastewater treatment plant, whose effluent is discharged directly into the River Tongyang after routine treatment. To prevent bias in the model results, the impact of this wastewater treatment plant’s discharge was excluded during the data processing phase. The assessment of the potential impact of agricultural non-point source pollution conducted in this study focuses mainly on areas dominated by agricultural activities and does not include the potential effects of non-agricultural pollution on the water environment, such as rural domestic sewage or surface runoff from residential areas.

2.3. Methods

This study focuses on the main crops in the River Tongyang Watershed and divides calculation units based on regional water resource management principles. By considering the differences in nitrogen and phosphorus loss coefficients under different land use types, as well as the temporal and spatial variations in soil, crops, and residue characteristics, the losses of agricultural non-point source pollutants within the watershed are calculated. The modeling approach mainly consists of three analytical modules: the hydrological response module converts meteorological inputs into the watershed’s hydrodynamic field; the pollution load and transfer module simulates the ‘source-sink’ transfer and transformation processes of nitrogen and phosphorus within the watershed; and the environmental capacity warning module performs reverse regulation based on water quality targets, providing a quantitative management basis for agricultural fertilization. Within this framework, the model can predict the maximum amount of fertilization the watershed can withstand on a given day based on future weather conditions, thereby achieving dynamic optimization of fertilization management. Considering that pollutant transfer is influenced by natural factors such as soil, terrain, and rainfall, this study selected four key factors: relative elevation, slope, rainfall, and land use type. These factors can finely characterize the spatial distribution of agricultural non-point source pollution within the watershed, providing a scientific basis for dynamic fertilization management and water quality risk warning. The selection of these factors is based on the characteristics of the River Tongyang Watershed and previous research in this field [19,20,21,22], as detailed in

Table 1. Factors affecting AGNPSP in the River Tongyang Watershed.

Index	Indicator Description
Rainfall	An increase in rainfall will affect pollutant wash-off and runoff transport.
Slope	Slope affects surface runoff velocity and soil erosion intensity.
Land use type	Regulation of pollutant source intensity and transport characteristics under different utilization methods.
Digital elevation Model (DEM)	The impact of terrain elevation on pollutant transport pathways and accumulation effects.

.

2.3.1. Hydrological Response

Hydrological processes serve as the carrier for pollutant migration, and accurately simulating the continuous process from rainfall to flow rate is the physical basis for the accuracy of early warning systems.

(a)

Rainfall-runoff process. This study uses the SCS-CN model [23] for secondary rainfall-runoff calculations. This model can effectively consider the impact of underlying surface conditions of the watershed on the runoff process, and its calculation formula is as follows:

$$Q=\left\{\begin{array}{c}{\displaystyle \frac{{\left(P-0.2S\right)}^2}{P+0.8S}}, P>0.2S\\ 0, P\le 0.2S\end{array}\right.$$

(1)

$$S={\displaystyle \frac{25400}{CN}}-254$$

(2)

where Q is the surface runoff depth (mm); P is the forecasted rainfall (mm); S is the maximum retention of the watershed (mm), determined by the Curve Number (CN).

The CN value is a key parameter of the model. Most of the farmland in the River Tongyang Watershed is rain-fed, with some fields used for rice cultivation in the spring. This study first classified the watershed’s soils into two hydrological groups, A and B, based on soil texture, and then assigned corresponding basic runoff coefficients to different plots using land use data, thereby accurately characterizing the spatial heterogeneity of the watershed’s runoff-generating capacity.

(b)

Hydrodynamic processes of river channels. River flow velocity is a key factor in determining the transport time of pollutants. During the dry period, hydrological measurements were conducted in the straight sections of the middle and upper reaches of the River Tongyang to obtain the flow velocity v₁ and river cross-sectional data under baseflow conditions. Combined with the riverbed slope S extracted from the DEM, the Manning roughness coefficient n for this river section was back-calculated using the Manning formula [24]. At the same time, the baseflow discharge Q_b was determined from measured data. In routine simulations, the river flow velocity on rainless days is taken as the baseflow velocity v₁ determined by this n value and Q_b.

When there is a forecast of rainfall, the contribution of rainfall runoff to river flow must be considered. In this study, the time-area method was applied for flow routing [25]. This method is based on the principle of equal travel time within the watershed. The procedure is as follows: First, the travel time for each point in the watershed was determined using a digital elevation model (DEM) and GIS hydrological analysis tools, including flow direction calculation, flow accumulation analysis, and estimation of flow path lengths to the outlet. Based on these travel times, the watershed was divided into several subareas of equal travel time. Next, the area of each subarea was calculated to establish the cumulative contribution area–travel time relationship. Finally, the runoff depth generated by the SCS-CN model for each subarea was routed to the outlet according to its travel time and linearly superimposed to construct the total flow hydrograph at the outlet. From this hydrograph, the peak flow increment ΔQ generated by the rainfall event at key sections of the River Tongyang main channel was estimated. The rainfall-induced flow increment was then added to the base flow to obtain the total flow Q_t at the river cross-section during the rainfall, as given by Equation (3).

Finally, the total flow Q_t is substituted into the Manning formula to estimate the flow velocity v, as shown in Equation (4):

$$Q_t=Q_b+∆Q$$

(3)

$$v={\displaystyle \frac{1}{n}}{R}^{{\displaystyle \frac{2}{3}}}{S}^{{\displaystyle \frac{1}{2}}}$$

(4)

2.3.2. Pollution Load Migration

(a)

Pollutant migration. This study is based on DEM data with a resolution of 30 m and uses GIS analysis tools to extract the river network of the River Tongyang Watershed and the flow path length D_i from each plot i to the lake inlet. Combined with the flow velocity v output from the hydrological module, the time t_i (days) required for pollutants to migrate from plot i to the lake inlet is calculated as follows:

$$t_i={\displaystyle \frac{D_i}{v\times 86400}}$$

(5)

(b)

Dynamic output coefficients and estimation of pollutant load. Pollutant output coefficients [26] are the key link connecting human activities to water pollution load. Extensive research has shown that output coefficients are not constant values; they strongly depend on rainfall intensity, prior soil moisture, farming practices, and the time interval between fertilization and rainfall events. To more scientifically characterize this dynamic process, a rainfall-adjusted dynamic output coefficient scheme is introduced. Since the migration and transformation processes of nitrogen and phosphorus are similar, the subsequent calculations and analyses in this paper will focus solely on total phosphorus concentration.

$$L_i=L_{base}\left(i\right)\times \left(1+R_f\right)$$

(6)

$$R_f=\left\{\begin{array}{c}a\times {\left(P-P_0\right)}^b,P>P_0\\ 0, P<P_0\end{array}\right.$$

(7)

where L_i is the total phosphorus transfer coefficient for plot i, reflecting the proportion of applied phosphorus fertilizer that enters surface runoff. L_base(i) represents the baseline total phosphorus output coefficient for plot i under standard conditions. Based on studies on non-point source pollution output coefficients in plain river networks, in this study, L_base(i) is taken as 0.18. This study comprehensively refers to the “National Waterbody Pollution Load Estimation Manual” and empirical research literature published in the agricultural regions of eastern China [27,28,29,30], assigning locally validated initial values to different land use types such as paddy fields and drylands. R_f is the rainfall-driven factor used to quantify the amplification effect of specific rainfall events on the output coefficient. P is the forecasted rainfall, P₀ is the rainfall threshold required to generate significant runoff (mm). Based on previous studies on runoff characteristics in the Lake Chaohu Watershed [31], P₀ is set to 5 mm in this study. Parameters a and b are empirical coefficients; after multiple experiments, this study selects a = 0.02 and b = 0.9, and the model was locally calibrated using historical monitoring data during the verification phase.

On this basis, calculate the mass of phosphorus m_i,ent from plot i that may enter the water body during this fertilization event:

$$m_{i,ent}=M_i\times L_i\times {10}^6$$

(8)

where M_i is the mass of phosphorus contained in the fertilizer applied to plot i (kg).

(c)

River channel migration and contribution to lake inlet concentration. During the migration of pollutants along the water flow, degradation occurs. To quantify this process, this study introduces the pollutant residual rate η_i (Equation (9)), which is based on a first-order kinetic model describing pollutant decay in the environmental model used in this study to quantify this effect [32]. Subsequently, Equation (10) calculates the contribution of the pollutant load from plot i to the total phosphorus concentration at the lake inlet. Finally, Equation (11) is used to calculate the total concentration increase generated by the n plots within the watershed.

$${\eta }_i=e^{-k{t}_i}$$

(9)

$$m_{i,out}=m_{i,ent}\times {\eta }_i$$

(10)

$$∆C_i={\displaystyle \frac{m_{i,out}}{V}}$$

(11)

$$∆C_{all}=\sum _{i=1}^n∆C_i=\sum _{i=1}^n{\displaystyle \frac{M_i\times L_i\times {10}^6\times {\eta }_i}{V}}$$

(12)

where η_i represents the proportion of the pollutant mass reaching the lake outlet after a migration time t_i. k is the comprehensive degradation coefficient of total phosphorus in the river water, which is taken as 0.17 in this study, based on research results under similar climatic and hydrological conditions [33,34,35]. V is the main water volume of the River Tongyang channel (L), estimated based on the river channel morphology.

2.3.3. Environmental Capacity Fertilization Prediction

(a)

Water environmental capacity is the maximum load of pollutants that a water body can accommodate while maintaining its functional objectives [36]. This study dynamically calculates the remaining environmental capacity, i.e., the allowable concentration increment, based on real-time water quality monitoring data at the River Tongyang’s lake inflow section, serving as a benchmark for fertilization warnings:

$$∆C_a=max\left(0,C_{st}-C_{bg}\right)$$

(13)

where ∆C_a is the remaining environmental capacity of the current river’s total phosphorus concentration (mg/L), C_st is the total phosphorus concentration limit for Class V water quality set for lakes and reservoirs according to the ‘Environmental Quality Standards for Surface Water’ of the People’s Republic of China (0.2 mg/L). C_bg is the real-time measured concentration at the automatic water quality monitoring station on the day the warning is issued (mg/L). The max function ensures that when the background concentration has already exceeded the standard, the allowable increment is zero, and no additional fertilization activities are recommended.

Convert the allowable concentration increment into the maximum mass of phosphorus permitted to enter the lake, and calculate the maximum mass of phosphorus allowed to enter the lake for the entire watershed.

$$M_{Permit}=∆C_a\times V$$

(14)

where M_Permit is the maximum allowable phosphorus load from all plots in the entire watershed reaching the lake inlet (kg).

(b)

M_Permit represents the mass of phosphorus that reaches the lake inlet after attenuation during river transport. Fertilizer phosphorus applied to farmland undergoes two major attenuations before reaching the lake inlet: first, surface runoff loss, as not all applied fertilizer is carried into rivers by runoff, a proportion described by the dynamic output coefficient L_i; second, riverine transport loss, where pollutants entering the river degrade during transport, with the proportion reaching the lake inlet indicated by the residual rate η_i. Therefore, it is necessary to calculate the maximum allowable total phosphorus fertilizer application M_TP based on the final permissible phosphorus mass at the lake outlet, M_Permit. The total mass of phosphorus fertilizer applied to all plots in the watershed, after accounting for output and migration attenuation, should equal the permissible mass M_Permit (Equation (15)), thereby determining the maximum allowable total phosphorus application (Equation (16)):

$${\sum }_{i=1}^n\left(M_{TP}\times {\displaystyle \frac{s_i}{\sum S_i}}\times L_i\times {\eta }_i\right)=M_{Permit}$$

(15)

$$M_{TP}={\displaystyle \frac{M_{Permit}}{{\sum }_{i=1}^n\left(\frac{s_i}{\sum S_i}\times L_i\times {\eta }_i\right)}}$$

(16)

where M_TP is the total mass of phosphorus fertilizer allowed for the watershed (kg). S_i is the area of plot i (ha).

(c)

The total phosphorus load allowed to enter the lake is contributed collectively by various plots within the watershed. To ensure fairness and practicability in quota allocation, this plan follows the principle of distribution according to the proportion of arable land area, allocating the maximum permissible fertilization amount for the entire watershed back to the plot level based on plot area. On this basis, it is further converted into the maximum recommended fertilization amount per mu that farmers can practically apply, and graded warnings are issued accordingly. This provides a scientific and implementable decision-making basis for the dynamic control and refined management of agricultural non-point source pollution in the River Tongyang watershed. Under the condition of satisfying environmental capacity constraints, the formulas for calculating the corresponding total fertilization amount and the fertilization amount for each plot are as follows:

$$M_F={\displaystyle \frac{M_{TP}}{\rho }}$$

(17)

$$M_{i,F}=M_{F\times }{\displaystyle \frac{S_i}{{\sum }_{i=1}^n{S}_i}}$$

(18)

where M_F is the total permissible amount of fertilizer for the watershed (kg), ρ is the mass fraction of phosphorus in the applied fertilizer, M_i,F is the total amount of fertilizer that can be applied to plot i (kg), and S_i is the area of plot i (ha).

(d)

To quantitatively analyze the impact of different plots on pollutant concentrations at lake inflows, this study introduces the concept of pollution potential. Here, pollution potential is defined as the potential contribution of the fertilizer applied per unit area of a plot to the pollutant concentration at the lake inflow, reflecting the relative intensity of a plot’s influence on downstream water quality changes. The pollution output of a plot primarily depends on the effect of its area, S. Under the same fertilization intensity and management conditions, a larger plot has a higher potential pollution source base. Secondly, the fertilizer applied to a plot must enter the river system through surface runoff, and its loss efficiency can be represented by a dynamic output coefficient, L_i, indicating the proportion of fertilizer per unit area that reaches the river. This coefficient comprehensively reflects the cumulative influence of soil type, terrain slope, land use, and rainfall conditions on pollutant loss. Once pollutants enter the river, not all of them can migrate to the lake inflow; during transport, they are affected by multiple processes, including dilution, adsorption, sedimentation, and biological uptake. For this reason, the pollutant retention rate, η_i, is used to characterize the transport efficiency of pollutants from the river to the lake inflow, reflecting the attenuation effects caused by spatial differences in plots. Therefore, the total pollution potential of a plot for lake inflow concentration is proportional to the continuous interplay of these three factors; the product of these factors represents the contribution of the plot to the total phosphorus concentration at the lake inflow. The significance of this product is that when the fertilization per unit area of plot i is one unit, the theoretical contribution to the total phosphorus concentration at the lake inflow is:

$${\omega }_i=S_i·L_i·{\eta }_i$$

(19)

Based on the results of pollution potential calculations, this paper classifies each plot within the watershed into four levels: high risk, relatively high risk, medium risk, and low risk, to conduct spatial risk differentiation analysis.

3. Results

3.1. Model Accuracy Verification

To ensure the reliability of the pollutant migration simulation in the early warning model, this study selected the 2024 period of concentrated fertilizer application by farmers as the validation period. Through field surveys, the habitual fertilization amounts per mu and timing in typical farmlands within the watershed were obtained, and combined with measured rainfall and runoff process data, serving as input parameters for the model. At the same time, automatic monitoring stations at the lake inflow recorded total phosphorus concentrations continuously during the same period, which were used to compare and validate the model simulation results. The modeled total phosphorus concentrations at the lake inflow were compared with measured data for simulations from April to May and September to October (Figure 3). Overall, the trends of the simulated values were highly consistent with the measured data. The model successfully captured the rapid increase in total phosphorus concentrations following major rainfall events and the subsequent attenuation characteristics. The timing of peak concentrations was generally consistent with observed lag features, indicating that the model can adequately reflect the dynamic response of nonpoint source pollution migration driven by rainfall runoff.

To quantitatively assess model accuracy, this study used statistical indicators such as the coefficient of determination R², root mean square error (RMSE), and percent bias (PBIAS) for validation. The calculation of these four indicators, respectively, reflects the goodness of fit, overall error level, and systematic bias of the model, with their formulas as follows:

$$R^2={\left[{\displaystyle \frac{\sum \left(y_{i,sim}-{\overline{y}}_{i,sim}\right)\left(y_{i,mea}-{\overline{y}}_{i,mea}\right)}{\sqrt{\sum {\left(y_{i,sim}-{\overline{y}}_{i,sim}\right)}^2\sum {\left(y_{i,mea}-{\overline{y}}_{i,mea}\right)}^2}}}\right]}^2$$

(20)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}\sum {\left(y_{i,sim}-y_{i,mea}\right)}^2}$$

(21)

$$PBIAS=100\times {\displaystyle \frac{\sum \left(y_{i,sim}-y_{i,mea}\right)}{\sum y_{i,mea}}}$$

(22)

where $y_{i,sim}$ represents the predicted value of total phosphorus concentration at the lake inlet, and $y_{i,mea}$ represents the actual value of the predicted total phosphorus concentration at the lake inlet.

The calculation results indicate that the model has an R² of 0.89, an RMSE of 0.019 mg/L, and a PBIAS of −2.8%. Based on these statistical results, the following conclusions can be drawn:

R² = 0.89 indicates a high degree of fit between the model and the observed data, showing good model performance. An R² value greater than 0.75 is considered a good standard for model fit [37], indicating that the model can well capture the variation trends in the data.

RMSE = 0.019 mg/L, and the calculated results yield a measured mean value of 0.099 mg/L. The RMSE relative to the measured mean is approximately 19%, indicating that the model prediction error is relatively small [38]. This result suggests that the model is able to capture the variation trend of total phosphorus concentration at the lake inlet well and has good predictive capability.

PBIAS = −2.8% indicates that the model’s systematic bias is small [39], and the overall bias is negative, showing that the model slightly underestimates the total phosphorus concentration. Considering the uncontrollability of fertilization timing and the omission of point source pollution, the model can still effectively reflect water quality changes within the error range.

Although the overall model fit is good, there are certain deviations between the simulated and observed values on specific dates. On some dates, the simulated values are slightly lower than the observed ones, while on others they are slightly higher. The main causes of these discrepancies may include: uncertainty in the timing of fertilization in certain fields, which leads to biases in the model input data; the model not explicitly considering point source emissions such as rural domestic sewage, which may result in underestimation of short-term peaks of dissolved phosphorus; and some rainfall events failing to generate significant runoff, causing the model to overestimate phosphorus concentrations in these cases. In summary, despite these local discrepancies, their impact on the overall simulation accuracy is relatively limited, and the model is able to reasonably reflect the temporal characteristics of total phosphorus transport in the River Tongyang Watershed under rainfall-runoff conditions, providing support for subsequent fertilization limit control and spatial analysis of pollution risks.

3.2. Spatial Variation in Fertilization Limits and Pollution Risks in River Watersheds

Based on the model accuracy validation, in order to further evaluate the application potential of this early-warning model in controlling agricultural non-point source pollution, this study systematically analyzed the spatial differentiation characteristics of fertilizer limits and pollution risks under different rainfall scenarios. By extracting the cultivated land area from land use classification data, the actual cultivated area of each plot was determined, and, combined with the model calculation results, the spatial distribution of plot-specific fertilizer limits under different rainfall conditions was obtained. Considering the differences in phosphorus content among various fertilizer products, to ensure a direct correspondence between fertilizer limits and water quality targets, this study used the mass of phosphorus in the fertilizer as a unified analysis indicator and assumed that all plots were fertilized within a similar time frame, calculating the total permissible fertilizer amount for the watershed under the constraint of environmental capacity. Field survey results showed that the phosphorus content in base fertilizer for local farmers’ wheat was about 50 kg/ha, while for rice cultivation, it was about 25 kg/ha.

The model’s dynamic response of fertilizer limits under different rainfall scenarios is shown in Figure 4 and

Table 2. Fertilizer quantity limits and recommendations based on meteorological information.

Rainfall Level	Total Phosphorus Concentration at Lake Inlet (mg/L)	Allowed Phosphorus Fertilization Amount (kg/ha)	Suggestions
No rain	0.10	≤50.0	Fertilization can proceed as originally planned.
Light rain	0.10	≤43.0	Fertilization can basically proceed as planned.
Light rain	0.15	30.5–34.5	Slightly reduce the amount of fertilizer applied.
Moderate rain	0.10	28.0–32.5	Slightly reduce the amount of fertilizer applied.
Moderate rain	0.15	16.5–21.0	Fertilizer application needs to be greatly reduced.
Heavy rain or above	0.10	0	Fertilization is not recommended during heavy rain.
Others	≥0.2	0	Fertilization prohibited due to excessive concentration.

, and the results reflect significant rainfall-driven characteristics. Figure 4 illustrates the spatial distribution of fertilizer limits for each plot under light rain (Figure 4a) and moderate rain (Figure 4b) scenarios, assuming a total phosphorus concentration at the lake inlet of 0.15 mg/L. As shown in the figure, the northern part of the watershed, due to concentrated farmland, larger areas, and relatively flat terrain, has a strong fertilizer carrying capacity and therefore receives a relatively high total amount of fertilizer. In contrast, the western mountainous terrain, with its large elevation variations, scattered farmland, and limited area, results in significantly lower corresponding fertilizer limits. Under the same environmental capacity constraints, changes in rainfall can clearly regulate the fertilizer limits of plots by affecting surface runoff and nutrient migration processes.

Table 2 is the fertilization warning table for agricultural non-point source pollution in the River Tongyang Watershed. The allowable phosphorus fertilization amounts under each scenario in the table are calculated using Formulas (13)–(17). For a given rainfall category and total phosphorus concentration at the lake inlet, the model first determines the corresponding rainfall input and background concentration, then simulates the transport of pollutants from the plot to the lake inlet, and ultimately back-calculates the maximum fertilization rate per unit area that meets the water quality targets at the lake inlet. The fertilization amounts presented in the table are given as ranges to reflect the variations in calculation results corresponding to the rainfall range for that category (e.g., light rain: 1–10 mm).

As shown in Table 2, when the actual pollutant concentration at the lake inlet is low and rainfall is minimal, the contribution of surface runoff is small, and pollutants mainly remain in the soil in an adsorbed form. In this case, the total fertilization amount that the watershed can tolerate is sufficient, and farmers can fertilize according to the original plan based on the needs of their crops. As rainfall increases or the actual pollutant concentration at the lake inlet rises, the runoff process intensifies, making it easier for pollutants to be transported into the river with runoff, requiring a proportional reduction in fertilization amounts. When rainfall is excessive, considering the actual farming and fertilization conditions, farmers are advised not to fertilize. If the background concentration is already close to or exceeds the standard, fertilization is strictly prohibited. This feature indicates that the model can adaptively adjust fertilization management strategies according to rainfall scenarios, demonstrating strong dynamic warning and control capabilities.

From the perspective of pollution risk, significant spatial differences are also evident between different areas of the watershed. According to the method for calculating pollution potential described in Formula (19), this paper categorizes the contribution of each plot to the total phosphorus concentration at the lake inlet into different risk levels, as shown in Figure 5. In the upper and middle reaches, the terrain is highly undulating, slopes are steep, soil permeability is poor, and runoff generation capacity is strong. Although the model assigns lower fertilizer limits, the pollutant output intensity per unit area is high, making the risk of agricultural non-point source pollution more prominent. In contrast, the downstream plain areas have gentle terrain, good infiltration conditions, and slower runoff convergence. They exhibit strong nitrogen and phosphorus adsorption and retention capacities, reflecting lower migration risks. However, plots closer to the lake inlet still face relatively high pollution risks. It is therefore apparent that the spatial distribution of fertilizer limits in the watershed does not completely align with the spatial distribution of pollution risks. This inconsistency reflects the comprehensive regulatory effects of topographical conditions, land use types, and hydrological processes on the formation and diffusion of agricultural non-point source pollution. Elevation differences determine runoff paths and convergence speeds, while soil texture and land use influence nutrient migration and retention processes. Thus, at the watershed scale, scientifically setting differentiated fertilizer limits helps to ensure agricultural productivity while achieving a dynamic balance in total pollution load and continuously improving water environmental quality.

The early warning model developed in this study not only reveals the spatial distribution patterns of fertilizer limits in the River Tongyang Watershed under area constraints but also reflects the significant differences in pollution risk among different geomorphic units. The model results can provide a scientific basis for controlling total fertilizer use and optimizing spatial distribution at the watershed scale, as well as offer technical support for the refined management of agricultural non-point source pollution.

4. Discussion

4.1. Selection of Pollutants and Setting of Environmental Factors

The model simulates using phosphorus as a representative pollutant during the construction and validation processes. The reason is that major agricultural pollutants such as nitrogen and phosphorus have similar migration and transformation mechanisms in surface runoff. To simplify the complexity of model calculations and result analysis while ensuring the representativeness of the simulation, this study uses phosphorus as the core index. Phosphorus not only has significant soil retention characteristics but also has a considerable impact on the eutrophication of water bodies [40]. Therefore, using phosphorus as a pollutant indicator can effectively reflect the key processes of agricultural non-point source pollution under conditions of limited data.

The risk of agricultural non-point source pollution is influenced by various natural environmental factors, so selecting appropriate environmental factors is key to accurately simulating non-point source pollution output. The basic premise of AGNPSP is that fertilizers not fully absorbed and utilized by crops migrate and accumulate in rivers through soil erosion, rainfall, and sediment transport [41]. Clearly, natural factors such as topography, precipitation, vegetation cover, and soil conditions have a significant impact on the AGNPSP process. Wang et al. [42] selected seven indicator factors, including elevation, slope, rainfall erosion force, and distance from water, in their study of the upper Yangtze River; Xie et al. [43] conducted a coupled analysis of agricultural non-point source pollution in the Han River Basin, selecting six indicator factors, such as the normalized difference vegetation index, land use, and soil erosion, to characterize the migration and transformation of pollutants. Based on these experiences, this study selected four key natural environmental factors in the River Tongyang Watershed: terrain elevation, slope, rainfall, and land use type. These factors not only reflect the control of terrain on runoff, erosion, and pollutant migration pathways, but also encompass the regulation of land use on pollution source intensity, thereby effectively enhancing the model’s ability to represent spatial heterogeneity. In addition, considering the availability and integrity of the data, the selected factors take into account both theoretical rationality and practical operability, providing a scientific basis for dynamic fertilization management and water quality risk warning.

4.2. Research Limitations

Although the early warning model developed in this study has achieved good results in fertilizer management and non-point source pollution risk assessment, there are still the following limitations that need to be further addressed in future work.

This study focuses on non-point source phosphorus pollution generated during agricultural production. To maintain the specificity of the indicator system and the clarity of the model structure, domestic sewage discharge, livestock and poultry farming wastewater, and other non-agricultural sources were excluded from the model. However, in practical applications, non-agricultural pollution may still contribute to certain periods and local areas. Because non-agricultural pollution was excluded, some observed concentration peaks may originate from these pollutants, which the model does not account for, potentially leading to underestimation at certain times. Therefore, this model is more suitable for watershed management scenarios dominated by agricultural loads, while its applicability is limited in watersheds with strong, multiple-source interference. Future research will combine monitoring and socioeconomic statistical data to quantitatively supplement non-agricultural sources in order to enhance the explanatory power and universality of the model.

The simplification of river hydrodynamic processes is also one of the factors affecting model accuracy. The River Tongyang has various small-scale disturbances, such as local siltation, bed undulations, and abrupt changes in cross-sections. In this study, due to the lack of certain hydrodynamic data, a simplified approach was adopted. Although this method can meet the needs for long-term trend assessment and management scenario simulations, it is still insufficient to accurately depict short-term runoff during rainfall events and the initial flushing effects of pollutants. This simplification, to some extent, limits the model’s ability for detailed simulation.

4.3. Applications and Challenges of Agricultural Fertilization Management

Effectively communicating fertilization recommendations to farmers in a timely manner and supervising their implementation is a significant challenge in the application of research findings. The implementation of optimized fertilization strategies depends not only on the accuracy of the models themselves but also on factors such as farmers’ acceptance, the penetration of digital tools, and the capacity of the grassroots agricultural extension system. Therefore, to enhance the practicability of model results in agricultural management, the interactive web platform currently under development in this study integrates the core model algorithms, weather forecasts, and field-level crop information to provide decision support to users. For farmers, the platform can generate personalized fertilization plans at the field level and deliver them via mobile terminals or push notifications, enabling farmers to conveniently access scientifically guided fertilization instructions. For agricultural management authorities, the platform offers a watershed supervision interface that can display real-time water quality monitoring data at lake inlets, risk zoning results, and pollution contributions from specific fields, helping management authorities identify high-risk areas, issue management recommendations, and organize targeted measures. Through these mechanisms, the research outcomes can establish effective communication channels among researchers, agricultural managers, and farmers, thereby promoting the practical implementation of agricultural fertilization management.

5. Conclusions

This study focuses on the River Tongyang Watershed in Lake Chaohu and develops a reverse-calculation early warning model for agricultural non-point source pollution based on water environmental capacity constraints. Centered on the total phosphorus control target for the inflow into the lake, the model treats water environmental capacity as a rigid constraint and integrates rainfall forecasts, surface runoff, and pollutant migration processes, achieving a shift from pollution outcome assessment to risk source management. The precise result is as follows:

(1)

This study established a hydrological response module and a pollution load and transport module to simulate the characteristics of total phosphorus output from agricultural non-point sources in the River Tongyang Watershed during rainfall events. A comparison between simulation results and monitoring data validated the model, achieving an R² of 0.81, an RMSE of 0.06 mg/L, and a PBIAS of −7.4%. The results indicate that the model can accurately capture the peak response of total phosphorus concentration after rainfall events and its attenuation process, demonstrating strong temporal prediction capability and providing a reliable basis for the development of fertilization warning systems.

(2)

By introducing the concept of pollution potential, the contribution of different plots to pollutant concentrations at the lake inlet was quantitatively assessed, and the AGNPSP risk in the River Tongyang Watershed was classified into four categories: high risk, relatively high risk, medium risk, and low risk. The high-risk and relatively high-risk areas account for 70% of the total watershed area, mainly located in the mid-to-upper reaches with steep slopes, low soil permeability, strong runoff capacity, and proximity to the lake inlet, where the pollutant output intensity per unit area is significantly higher. In the mid-to-lower reaches, the plains are flat with good infiltration conditions and slower flow convergence, possessing strong nitrogen and phosphorus adsorption and retention capabilities, and thus lower migration risk. Overall, the agricultural non-point source pollution pressure in the River Tongyang Watershed is considerable, highlighting the need for further research on pollution mechanisms and control measures.

(3)

A calculation method for the fertilization limit that can be dynamically updated according to weather forecasts has been established. Based on meteorological data, the model can calculate the fertilization limit needed to meet water quality standards and refine control indicators to the plot level through a spatial allocation algorithm, achieving a shift from post-event evaluation to preemptive prevention. This study transforms abstract water quality standards into actionable fertilizer limits per unit area and, by integrating rainfall forecasts, enables the model to anticipate fertilization restrictions under different rainfall conditions. This provides a scientific basis for formulating dynamic and zoned management measures, supporting fertilizer scheduling at the watershed level, water quality protection, and the sustainable development of agriculture.

In summary, although the study has some limitations, these shortcomings point to valuable directions for future research. In addition, due to the lack and insufficiency of some monitoring data, data will be continuously collected and updated in subsequent studies to conduct a more detailed analysis of the risk patterns of AGNPSP and the migration of pollutants, further improving the accuracy of the predictive data.