Uncertainty and multiple objective calibration in regional water balance modelling: case study in 320 Austrian catchments

We examine the value of additional information in multiple objective calibration in terms of model performance and parameter uncertainty. We calibrate and validate a semi‐distributed conceptual catchment model for two 11‐year periods in 320 Austrian catchments and test three approaches of parameter calibration: (a) traditional single objective calibration (SINGLE) on daily runoff; (b) multiple objective calibration (MULTI) using daily runoff and snow cover data; (c) multiple objective calibration (APRIORI) that incorporates an a priori expert guess about the parameter distribution as additional information to runoff and snow cover data. Results indicate that the MULTI approach performs slightly poorer than the SINGLE approach in terms of runoff simulations, but significantly better in terms of snow cover simulations. The APRIORI approach is essentially as good as the SINGLE approach in terms of runoff simulations but is slightly poorer than the MULTI approach in terms of snow cover simulations. An analysis of the parameter uncertainty indicates that the MULTI approach significantly decreases the uncertainty of the model parameters related to snow processes but does not decrease the uncertainty of other model parameters as compared to the SINGLE case. The APRIORI approach tends to decrease the uncertainty of all model parameters as compared to the SINGLE case. Copyright © 2006 John Wiley & Sons, Ltd.     

Integrating tracer experiments with modeling to assess runoff processes and water transit times

Representing runoff process complexity in a simple model structure remains a challenge in hydrology. We present an integrated approach to investigate runoff processes using a hillslope tracer experiment and modeling exercise to explore model parameterization, process representation, and transit times. A spatially-explicit model constrained by soil hydrologic properties, runoff, and applied tracer data was used to identify the dominant processes necessary to explain both water and solute flux from a steep hillslope. The tracer data allowed for the rejection of model parameter sets based on the calibration to runoff data alone, thus reducing model uncertainty. The additional calibration to tracer data, improved parameter identifiability and provided further insight to process controls on hillslope-scale water and solute flux. Transit time distributions developed using the model provided further insight to model structure such as subsurface volume, mixing assumptions, and the water table dynamics. Combining field experiments with the modeling exercise may lead to a more comprehensive assessment of runoff process representation in models.     

The comparative study of multi‐site uncertainty evaluation method based on SWAT model

With the recent development of distributed hydrological models, the use of multi‐site observed data to evaluate model performance is becoming more common. Distributed hydrological model have many advantages, and at the same time, it also faces the challenge to calibrate over‐do parameters. As a typical distributed hydrological model, problems also exist in Soil and Water Assessment Tool (SWAT) parameter calibration. In the paper, four different uncertainty approaches – Particle Swarm Optimization (PSO) techniques, Generalized Likelihood Uncertainty Estimation (GLUE), Sequential Uncertainty Fitting algorithm (SUFI‐2) and Parameter Solution (PARASOL) – are taken to a comparative study with the SWAT model applied in Peace River Basin, central Florida. In our study, the observed river discharge data used in SWAT model calibration were collected from the three gauging stations at the main tributary of the Peace River. Behind these approaches, there is a shared philosophy; all methods seek out many parameter set to fit the uncertainties due to the non‐uniqueness in model parameter evaluation. On the basis of the statistical results of four uncertainty methods, difficulty level of each method, the number of runs and theoretical basis, the reasons that affected the accuracy of simulation were analysed and compared. Furthermore, for the four uncertainty method with SWAT model in the study area, the pairwise correlation between parameters and the distributions of model fit summary statistics computed from the sampling over the behavioural parameter and the entire model calibration parameter feasible spaces were identified and examined. It provided additional insight into the relative identifiability of the four uncertainty methods Copyright © 2014 John Wiley & Sons, Ltd.     

Uncertainty in water resource model parameters used for climate change impact assessment

Despite their acknowledged limitations, lumped conceptual models continue to be used widely for climate‐change impact assessments. Therefore, it is important to understand the relative magnitude of uncertainties in water resource projections arising from the choice of model calibration period, model structure, and non‐uniqueness of model parameter sets. In addition, external sources of uncertainty linked to choice of emission scenario, climate model ensemble member, downscaling technique(s), and so on, should be acknowledged. To this end, the CATCHMOD conceptual water balance model was used to project changes in daily flows for the River Thames at Kingston using parameter sets derived from different subsets of training data, including the full record. Monte Carlo sampling was also used to explore parameter stability and identifiability in the context of historic climate variability. Parameters reflecting rainfall acceptance at the soil surface in simpler model structures were found to be highly sensitive to the training period, implying that climatic variability does lead to variability in the hydrologic behaviour of the Thames basin. Non‐uniqueness of parameters for more complex model structures results in relatively small variations in projected annual mean flow quantiles for different training periods compared with the choice of emission scenario. However, this was not the case for subannual flow statistics, where uncertainty in flow changes due to equifinality was higher in winter than summer, and comparable in magnitude to the uncertainty of the emission scenario. Therefore, it is recommended that climate‐change impact assessments using conceptual water balance models should routinely undertake sensitivity analyses to quantify uncertainties due to parameter instability, identifiability and non‐uniqueness. Copyright © 2005 John Wiley & Sons, Ltd.     

Application of a data‐based mechanistic modelling (DBM) approach for predicting runoff generation in semi‐arid regions

This paper addresses the application of a data‐based mechanistic (DBM) modelling approach using transfer function models (TFMs) with non‐linear rainfall filtering to predict runoff generation from a semi‐arid catchment (795 km²) in Tanzania. With DBM modelling, time series of rainfall and streamflow were allowed to suggest an appropriate model structure compatible with the data available. The model structures were evaluated by looking at how well the model fitted the data, and how well the parameters of the model were estimated. The results indicated that a parallel model structure is appropriate with a proportion of the runoff being routed through a fast flow pathway and the remainder through a slow flow pathway. Finally, the study employed a Generalized Likelihood Uncertainty Estimation (GLUE) methodology to evaluate the parameter sensitivity and predictive uncertainty based on the feasible parameter ranges chosen from the initial analysis of recession curves and calibration of the TFM. Results showed that parameters that control the slow flow pathway are relatively more sensitive than those that control the fast flow pathway of the hydrograph. Within the GLUE framework, it was found that multiple acceptable parameter sets give a range of predictions. This was found to be an advantage, since it allows the possibility of assessing the uncertainty in predictions as conditioned on the calibration data and then using that uncertainty as part of the decision‐making process arising from any rainfall‐runoff modelling project. Copyright © 2001 John Wiley & Sons, Ltd.     

Towards a simple dynamic process conceptualization in rainfall–runoff models using multi‐criteria calibration and tracers in temperate,upland catchments

Empirically based understanding of streamflow generation dynamics in a montane headwater catchment formed the basis for the development of simple, low‐parameterized, rainfall–runoff models. This study was based in the Girnock catchment in the Cairngorm Mountains of Scotland, where runoff generation is dominated by overland flow from peaty soils in valley bottom areas that are characterized by dynamic expansion and contraction of saturation zones. A stepwise procedure was used to select the level of model complexity that could be supported by field data. This facilitated the assessment of the way the dynamic process representation improved model performance. Model performance was evaluated using a multi‐criteria calibration procedure which applied a time series of hydrochemical tracers as an additional objective function. Flow simulations comparing a static against the dynamic saturation area model (SAM) substantially improved several evaluation criteria. Multi‐criteria evaluation using ensembles of performance measures provided a much more comprehensive assessment of the model performance than single efficiency statistics, which alone, could be misleading. Simulation of conservative source area tracers (Gran alkalinity) as part of the calibration procedure showed that a simple two‐storage model is the minimum complexity needed to capture the dominant processes governing catchment response. Additionally, calibration was improved by the integration of tracers into the flow model, which constrained model uncertainty and improved the hydrodynamics of simulations in a way that plausibly captured the contribution of different source areas to streamflow. This approach contributes to the quest for low‐parameter models that can achieve process‐based simulation of hydrological response. Copyright © 2009 John Wiley & Sons, Ltd.     

A GIS‐based variable source area hydrology model

Effective control of nonpoint source pollution from contaminants transported by runoff requires information about the source areas of surface runoff. Variable source hydrology is widely recognized by hydrologists, yet few methods exist for identifying the saturated areas that generate most runoff in humid regions. The Soil Moisture Routing model is a daily water balance model that simulates the hydrology for watersheds with shallow sloping soils. The model combines elevation, soil, and land use data within the geographic information system GRASS, and predicts the spatial distribution of soil moisture, evapotranspiration, saturation‐excess overland flow (i.e., surface runoff), and interflow throughout a watershed. The model was applied to a 170 hectare watershed in the Catskills region of New York State and observed stream flow hydrographs and soil moisture measurements were compared to model predictions. Stream flow prediction during non‐winter periods generally agreed with measured flow resulting in an average r² of 0·73, a standard error of 0·01 m³/s, and an average Nash‐Sutcliffe efficiency R² of 0·62. Soil moisture predictions showed trends similar to observations with errors on the order of the standard error of measurements. The model results were most accurate for non‐winter conditions. The model is currently used for making management decisions for reducing non‐point source pollution from manure spread fields in the Catskill watersheds which supply New York City's drinking water. Copyright © 1999 John Wiley & Sons, Ltd.     

Evaluation of storage–discharge relationships and recession analysis-based distributed hourly runoff simulation in large-scale,mountainous and snow-influenced catchment

ABSTRACT

Evaluation of a recession-based “top-down” model for distributed hourly runoff simulation in macroscale mountainous catchments is rare in the literature. We evaluated such a model for a 3090 km² boreal catchment and its internal sub-catchments. The main research question is how the model performs when parameters are either estimated from streamflow recession or obtained by calibration. The model reproduced observed streamflow hydrographs (Nash-Sutcliffe efficiency up to 0.83) and flow duration curves. Transferability of parameters to the sub-catchments validates the performance of the model, and indicates an opportunity for prediction in ungauged sites. However, the cases of parameter estimation and calibration excluding the effects of runoff routing underestimate peak flows. The lower end of the recession and the minimum length of recession segments included are the main sources of uncertainty for parameter estimation. Despite the small number of calibrated parameters, the model is susceptible to parameter uncertainty and identifiability problems.

EDITOR D. Koutsoyiannis; ASSOCIATE EDITOR A. Carsteanu     

Discharge and nutrient uncertainty: implications for nutrient flux estimation in small streams

The measurement of discharge is fundamental in nutrient load estimation. Because of our ability to monitor discharge routinely, it is generally assumed that the associated uncertainty is low. This paper challenges this preconception, arguing that discharge uncertainty should be explicitly taken into account to produce robust statistical analyses. In many studies, paired discharge and chemical datasets are used to calculate ‘true’ loads and used as the benchmark to compare with other load estimates. This paper uses two years of high frequency (daily and sub‐hourly) discharge and nutrient concentration data (nitrate‐N and total phosphorus (TP)) collected at four field sites as part of the Hampshire Avon Demonstration Test Catchment (DTC) programme. A framework for estimating observational nutrient load uncertainty was used which combined a flexible non‐parametric approach to characterising discharge uncertainty, with error modelling that allowed the incorporation of errors which were heteroscedastic and temporally correlated. The results showed that the stage–discharge relationships were non‐stationary, and observational uncertainties from ±2 to 25% were recorded when the velocity–area method was used. The variability in nutrient load estimates ranged from 1.1 to 9.9% for nitrate‐N and from 3.3 to 10% for TP when daily laboratory data were used, rising to a maximum of 9% for nitrate‐N and 83% for TP when the sensor data were used. However, the sensor data provided a better representation of the ‘true’ load as storm events are better represented temporally, posing the question: is it more beneficial to have high frequency, lower precision data or lower frequency but higher precision data streams to estimate nutrient flux responses in headwater catchments? Copyright © 2015 John Wiley & Sons, Ltd.     

Improving flow forecasting by error correction modelling in altered catchment conditions

Despite human is an increasingly significant component of the hydrologic cycle in many river basins, most hydrologic models are still developed to accurately reproduce the natural processes and ignore the effect of human activities on the watershed response. This results in non‐stationary model forecast errors and poor predicting performance every time these models are used in non‐pristine watersheds. In the last decade, the representation of human activities in hydrological models has been extensively studied. However, mathematical models integrating the human and the natural dimension are not very common in hydrological applications and nearly unknown in the day‐to‐day practice. In this paper, we propose a new simple data‐driven flow forecast correction method that can be used to simultaneously tackle forecast errors from structural, parameter and input uncertainty, and errors that arise from neglecting human‐induced alterations in conceptual rainfall–runoff models. The correction system is composed of two layers: (i) a classification system that, based on the current flow condition, detects whether the source of error is natural or human induced and (ii) a set of error correction models that are alternatively activated, each tailored to the specific source of errors. As a case study, we consider the highly anthropized Aniene river basin in Italy, where a flow forecasting system is being established to support the operation of a hydropower dam. Results show that, even by using very basic methods, namely if‐then classification rules and linear correction models, the proposed methodology considerably improves the forecasting capability of the original hydrological model. Copyright © 2013 John Wiley & Sons, Ltd.     

Attributes and uncertainty of dissolved zinc data from a mined catchment

This study evaluated the attributes and uncertainty of non‐point source pollution data derived from synoptic surveys in a catchment affected by inactive metal mines in order to help to identify and select appropriate methods for data analysis/reporting and information use. Dissolved zinc data from the Upper Animas River Basin, Colorado, USA, were the focus of the study. Zinc was evaluated because concentrations were highest relative to national water quality criteria for brown trout, and zinc had the greatest frequency of criteria exceedances compared with other metals. Data attributes evaluated included measurement and model error, sample size, non‐normality, seasonality and uncertainty. The average measurement errors for discharges, concentrations and loadings were 0·15, 0·1 and 0·18, respectively. The 90 and 95% coefficients of confidence intervals for mean concentrations based on a sample size of four were 0·48 and 0·65, respectively, and ranged between 0·15 and 0·23 for sample sizes greater than 40. Aggregation of data from multiple stations decreased the confidence intervals significantly, but additional aggregation of all data increased them as a result of increasing spatial variability. Unit area loading data were approximately log‐normal. Concentration data were right‐skewed but not log‐normal. Differences in median concentrations were appreciable between snowmelt and both storm flow and baseflow, but not between storm flow and baseflow. Differences in unit area loadings between all flow events were large. It was determined that the average concentration and unit area loading values should be estimated for each flow event because of significant seasonality. Time weighted values generally should be computed if annual information is required. The confidence in average concentrations and unit area loadings is dependent on the computation method used. Both concentrations and loadings can be significantly underestimated on an annual basis when using data from synoptic surveys if the first flush of contaminants during the initial snowmelt runoff period is not sampled. The ambient standard for dissolved zinc for all events was estimated as 1600 μg l⁻¹ using the 85th percentile of observed concentration data, with a 90% confidence interval width of 200 μg l⁻¹. Copyright © 1999 John Wiley & Sons, Ltd.     

Catchment‐scale sensitivity and uncertainty in water quality modelling

In this paper, we assess the performance of the catchment model SIMulated CATchment model (SIMCAT), to predict nitrate and soluble reactive phosphorus concentrations against four monitoring regimes with different spatial and temporal sampling frequencies. The Generalised Likelihood Uncertainty Estimation (GLUE) uncertainty framework is used, along with a general sensitivity analysis to understand relative parameter sensitivity. Improvements to model calibration are explored by introducing more detailed process representation using the Integrated Catchments model (INCA) water quality model, driven by the European hydrological predictions for the environment model. The results show how targeted sampling of headwater watercourses upstream of point discharges is essential for calibrating diffuse loads and can exert a strong influence on the whole‐catchment model performance. Further downstream, if the point discharges and loads are accurately represented, then the improvement in the catchment‐scale model performance is relatively small as more calibration points are added or frequency is increased. The higher‐order, dynamic model integrated catchments model of phosphorus dynamics, which incorporates sediment and biotic interaction, resulted in improved whole‐catchment performance over SIMCAT, although there are still large epistemic uncertainties from land‐phase export coefficients and runoff. However, the very large sampling errors in routine monitoring make it difficult to invest confidence in the modelling, especially because we know phosphorous transport to be very episodic and driven by high flow conditions for which there are few samples. The environmental modelling community seems to have been stuck in this position for some time, and whilst it is useful to use an uncertainty framework to highlight these issues, it has not widely been adopted, perhaps because there is no clear mechanism to allow uncertainties to influence investment decisions. This raises the question as to whether it might better place a cost on uncertainty and use this to drive more data collection or improved models, before making investment decisions concerning, for example, mitigation strategies. Copyright © 2016 John Wiley & Sons, Ltd.     

Uncertainty estimates for surface runoff models

A lower bound for variance in surface runoff modelling estimates is advanced. The bound is derived using a linear unit hydrograph approach which utilizes a discretization of the catchment into an arbitrary number of subareas, a linear routing technique for channel flow effects, a variable effective rainfall distribution over the catchment, and calibration parameter distributions developed in correlating rainfall-runoff data by the model. The uncertainty bound reflects the dominating influence of the unknown rainfall distribution over the catchment and is expressed as a distribution function that can be reduced only by supplying additional rainfall-runoff data. It is recommended that this uncertainty distribution in modelling results be included in flood control design studies in order to incorporate a prescribed level of confidence in flood protection facilities.     

SWAT-based runoff and sediment simulation in a small watershed,the loessial hilly-gullied region of China:capabilities and challenges

Model calibration and validation are necessary before applying it for scenario assessment and watershed management.This study presented the methodology of evaluating Soil and Water Assessment Tool(SWAT) and tested the feasibility of SWAT on runoff and sediment load simulation in the Zhifanggou watershed located in hilly-gullied region of China.Daily runoff and sediment event data from 1998-2008 were used in this study;data from 1998-2003 were used for calibration and 2004-2008 for validation.The evaluation statistics for the daily runoff simulation showed that the model results were acceptable,but the model underestimated the runoff for high-flow events.For sediment load simulation,the SWAT performed well in capturing the trend of sediment load,while the model tended to underestimate sediment load during both the calibration and validation periods. The disparity between observed and simulated data most likely resulted from limitations of the existing SCS-CN and MUSLE methods in the model.This study indicated that the modification of SWAT components is needed to take rainfall intensity and its duration into account to enhance the model performance on peak flow and sediment load simulation during heavy rainfall season.     

Sensitivity and identifiability of stream flow generation parameters of the SWAT model

Implementation of sensitivity analysis (SA) procedures is helpful in calibration of models and also for their transposition to different watersheds. The reported studies on SA of Soil and Water Assessment Tool (SWAT) model were mostly focused on identifying parameters for pruning or modifying during the calibration process. This paper presents a sensitivity and identifiability analysis of model parameters that influence stream flow generation in SWAT. The analysis was focused on evaluating the sensitivity of the parameters in different climatic settings, temporal scales and flow regimes. The global sensitivity analysis (GSA) technique based on classical decomposition of variance, Sobol', was employed in this study. The results of the study indicate that modeled stream flow show varying sensitivity to parameters in different climatic settings. The results also suggest that the identifiability of a parameter for a given watershed is a major concern in calibrating the model for the specific watershed, as it might lead to equifinality of parameters. The SWAT model parameters show varying sensitivity in different years of simulation suggesting the requirement for dynamic updation of parameters during the simulation. The sensitivity of parameters during various flow regimes (low, medium and high flow) is also found to be uneven, which suggests the significance of a multi‐criteria approach for the calibration of models. Copyright © 2010 John Wiley & Sons, Ltd.     

The identifiability of parameters in a water quality model of the Biebrza River, Poland

The identifiability of model parameters of a steady state water quality model of the Biebrza River and the resulting variation in model results was examined by applying the Monte Carlo method which combines calibration, identifiability analysis, uncertainty analysis, and sensitivity analysis. The water quality model simulates the steady state concentration profiles of chloride, phosphate, ammonium, and nitrate as a function of distance along a river. The water quality model with the best combination of parameter values simulates the observed concentrations very well. However, the range of possible modelled concentrations obtained for other more or less equally eligible combinations of parameter values is rather wide. This range in model outcomes reflects possible errors in the model parameters. Discrepancies between the range in model outcomes and the validation data set are only caused by errors in model structure, or (measurement) errors in boundary conditions or input variables. In this sense the validation procedure is a test of model capability, where the effects of calibration errors are filtered out. It is concluded that, despite some slight deviations between model outcome and observations, the model is successful in simulating the spatial pattern of nutrient concentrations in the Biebrza River.     

Input data resolution‐induced uncertainty in watershed modelling

The temporal‐spatial resolution of input data‐induced uncertainty in a watershed‐based water quality model, Hydrologic Simulation Program‐FORTRAN (HSPF), is investigated in this study. The temporal resolution‐induced uncertainty is described using the coefficient of variation (CV). The CV is found to decrease with decreasing temporal resolution and follow a log‐normal relation with time interval for temperature data while it exhibits a power‐law relation for rainfall data. The temporal‐scale uncertainties in the temperature and rainfall data follow a general extreme value distribution and a Weibull distribution, respectively. The Nash‐Sutcliffe coefficient (NSC) is employed to represent the spatial resolution induced uncertainty. The spatial resolution uncertainty in the dissolved oxygen and nitrate‐nitrogen concentrations simulated using HSPF is observed to follow a general extreme value distribution and a log‐normal distribution, respectively. The probability density functions (PDF) provide new insights into the effect of temporal‐scale and spatial resolution of input data on uncertainties involved in watershed modelling and total maximum daily load calculations. This study exhibits non‐symmetric distributions of uncertainty in water quality modelling, which simplify weather and water quality monitoring and reducing the cost involved in flow and water quality monitoring. Copyright © 2011 John Wiley & Sons, Ltd.     

Distributed discharge and sediment concentration predictions in the sub‐humid Ethiopian highlands: the Debre Mawi watershed

Experimental research in the Ethiopian highlands found that saturation excess induced runoff and erosion are common in the sub‐humid conditions. Because most erosion simulation models applied in the highlands are based on infiltration excess, we, as an alternative, developed the Parameter Efficient Distributed (PED) model, which can simulate water and sediment fluxes in landscapes with saturation excess runoff. The PED model has previously only been tested at the outlet of a watershed and not for distributed runoff and sediment concentration within the watershed. In this study, we compare the distributed storm runoff and sediment concentration of the PED model against collected data in the 95‐ha Debre Mawi watershed and three of its nested sub‐watersheds for the 2010 and 2011 rainy seasons. In the PED model framework, the hydrology of the watershed is divided between infiltrating and runoff zones, with erosion only taking place from two surface runoff zones. Daily storm runoff and sediment concentration values, ranging from 0.5 to over 30 mm and from 0.1 to 35 g l^?1, respectively, were well simulated. The Nash Sutcliffe efficiency values for the daily storm runoff for outlet and sub‐watersheds ranged from 0.66 to 0.82, and the Nash–Sutcliffe efficiency for daily sediment concentrations were greater than 0.78. Furthermore, the model uses realistic fractional areas for surface and subsurface flow contributions, for example between saturated areas (15%), degraded areas (30%) and permeable areas (55%) at the main outlet, while close similarity was found for the remaining hydrology and erosion parameter values. One exception occurred for the distinctly greater transport limited parameter at the actively gullying lower part of the watershed. The results suggest that the model based on saturation excess provides a good representation of the observed spatially distributed runoff and sediment concentrations within a watershed by modelling the bottom lands (as opposed to the uplands) as the dominant contributor of the runoff and sediment load. Copyright © 2014 John Wiley & Sons, Ltd.     

设施农业生产区降雨径流和氮磷输出特征及模拟——以滇池东岸花卉大棚种植区为例

为掌握滇池流域花卉大棚种植区的非点源污染特征,提高和改善滇池水环境质量,本研究选取呈贡县斗南村花卉大棚种植区作为研究对象,在实测降雨径流数据的基础上,通过建立Storm Water Management Model模型分别对全年连续降雨条件下和典型设计降雨条件下的降雨径流水质、水量进行了模拟.研究结果表明:1)模型的流量、化学需氧量(COD_(Cr))、悬浮物(SS)、总氮(TN)和总磷(TP)的Nash-Sutcliffe效率系数分别为0.858、0.835、0.803、0.712和0.752,能够较好地模拟研究区域的水质、水量变化.2)研究区域的平均径流系数为0.59,CODCr、SS、TN和TP的单位面积负荷率分别为118.34、82.90、54.64和5.46 kg/(hm~2·a),TN和TP是主要控制的污染物.3)各污染物浓度峰值的出现时间均早于流量峰值出现的时间,因此对滇池东岸花卉大棚种植区应进行污染物尤其是TP、TN浓度与流量错峰控制.     

Artificial neural network modelling of concentrations of nitrogen,phosphorus and dissolved oxygen in a non‐point source polluted river in Zhejiang Province,southeast China

A back‐propagation algorithm neural network (BPNN) was developed to synchronously simulate concentrations of total nitrogen (TN), total phosphorus (TP) and dissolved oxygen (DO) in response to agricultural non‐point source pollution (AGNPS) for any month and location in the Changle River, southeast China. Monthly river flow, water temperature, flow travel time, rainfall and upstream TN, TP and DO concentrations were selected as initial inputs of the BPNN through coupling correlation analysis and quadratic polynomial stepwise regression analysis for the outputs, i.e. downstream TN, TP and DO concentrations. The input variables and number of hidden nodes of the BPNN were then optimized using a combination of growing and pruning methods. The final structure of the BPNN was determined from simulated data based on experimental data for both the training and validation phases. The predicted values obtained using a BPNN consisting of the seven initial input variables (described above), one hidden layer with four nodes and three output variables matched well with observed values. The model indicated that decreasing upstream input concentrations during the dry season and control of NPS along the reach during average and flood seasons may be an effective way to improve Changle River water quality. If the necessary water quality and hydrology data are available, the methodology developed here can easily be applied to other case studies. The BPNN model is an easy‐to‐use modelling tool for managers to obtain rapid preliminary identification of spatiotemporal water quality variations in response to natural and artificial modifications of an agricultural drainage river. Copyright © 2009 John Wiley & Sons, Ltd.