Three rural catchments with different temperate climate, topography and farm types were monitored at high temporal resolution as part of the UK Demonstration Test Catchments (DTC) programme (Lloyd et al., 2016a, b; Outram et al., 2014; McGonigle et al., 2014). These were Newby Beck at Newby, Eden catchment, Cumbria (54.59∘ N, 2.62∘ W; 12.5 km2); Blackwater at Park Farm, Wensum catchment, Norfolk (52.78∘ N, 1.15∘ E; 19.7 km2); Wylye at Brixton Deverill, Avon catchment, Hampshire (51.16∘ N, 2.19∘ W; 50.2 km2) (Fig. 1). Further details of these catchments are available in Table S1 in the Supplement.

Rainfall was measured at 15 min resolution at three sites in each of the Newby Beck and Blackwater catchments (Outram et al., 2014; Perks et al., 2015) and summed to give hourly totals. The hourly totals from the different rain gauges were combined by areal weighting to give an hourly time series for the catchment. For the Wylye catchment, only daily rainfall was available for sites within the catchment, so raw tipping bucket data were obtained for several sites outside the catchment and analysed to produce an hourly time series which was considered most representative of the rainfall in the catchment. Further details of the rainfall analysis for the Wylye catchment are given in Sect. S1 in the Supplement.

River water level was measured at 15 min resolution in the three catchments, with rating curves developed for discharge estimation (Outram et al., 2014; Perks et al., 2015; Lloyd et al., 2016b). TP concentration was determined in situ at 30 min intervals with a Hach Lange combined Sigmatax sampling module and Phosphax analyser using acid digestion and colorimetry (Jordan et al., 2007, 2013; Perks et al., 2015). Total P loads for each hour were determined by multiplying discharge (averaged to 30 min resolution) by TP concentration for each 30 min and summing to give hourly totals: TP load(t)=k∑jQjCj, where TP load(t) is the load (kg) exported during the hourly timestep which ends at time t, Qj is the discharge observations (m3 s-1) within the hourly timestep, Cj is the corresponding TP concentration observations (mg L-1) within the hourly timestep, and k is a constant (= 3.6) for conversion of units to give load in kilograms. Visual inspection of the data indicated that aggregation of the data from 15 or 30 min resolution to hourly did not result in a significant loss of information. This would not be the case for very small catchments or those where the dynamics being investigated were very fast. Calculation of the load according to Eq. (1) assumes that the TP is well mixed in the water and that the Hach Lange sampler is taking a representative sample. It also assumes that the rating curve is appropriate over the full range of stage recordings made, and that the relationship between stage and discharge is stationary. Total phosphorus load, rather than concentration, was modelled because water utility companies are concerned about the total load which may have to be removed and because both water flow and load are fluxes, so comparisons between the two are easier to interpret directly than for concentration, which is a state rather than a flux (Jones et al., 2014).

Transfer function models relating the input (here, a time series of rainfall, R) to the output (here, a time series of either discharge, Q, or phosphorus load, TP load) were identified using continuous-time models (Young and Garnier, 2006) where possible, or in cases where data were missing or identification was difficult, with discrete time models (Young, 2003), the estimation of which handles missing data more robustly. Continuous-time models are more numerically robust and have a direct interpretation as systems of differential equations (Young, 2011). Models were identified using the RIVCBJ identification algorithm (refined instrumental variable continuous-time Box–Jenkins identification, for continuous-time models), or RIVBJ identification (refined instrumental variable Box–Jenkins identification, for discrete-time models) that are part of the CAPTAIN toolbox (Taylor et al., 2007) for MATLAB^®.

The identification algorithm always includes a noise model; by default this assumes normally distributed, uncorrelated errors, but an auto-regressive moving average (ARMA) structure can be specified. The Gaussian noise model still results in asymptotically unbiased parameter estimates, but not necessarily the most statistically efficient (close to minimum variance) (Taylor et al., 2007). In this study, models up to third order were considered initially, but higher order models showed no advantage, so only models up to second order were considered in subsequent evaluations. Full models (input–output (I-O) plus ARMA structured residual noise) were assessed initially and overall they did not produce better results in all cases; therefore, in order to keep a consistent approach for all catchments, structured noise models were not specified in later model identification. In addition, transfer function models with a structured noise component generally do not improve longer-term predictions of processes which are I-O dominated. The residuals structure was not strong enough for a structured noise model to improve the model fit consistently. If there was a strong structure in the residuals, it would suggest that something was being missed in the DBM system representation. The time delay constants were estimated from the data at the same time as the model structures.

Continuous-time and discrete-time model structures are described below (from Ockenden et al., 2017). The parameter estimates in both continuous-time models and discrete-time models are formulaically related (Table S3).

A second-order discrete linear transfer function, denoted by [2, 2, δ], takes the following form: y(t)=b1+b2z-11+a1z-1+a2z-2ut-δ+ξt, where y(t) is model output at time t, u(t) is model input, and z-1 is the backwards step operator, i.e. z-1y(t)=y(t-1); b1, b2, a1, a2 are parameters determined during model identification, δ is the number of time steps of pure time delay and ξt represents the uncertainty arising from a combination of measurement noise, other unmeasured inputs and modelling error. For a physical interpretation, second-order models were only accepted it they could be decomposed by partial fraction expansion into two first-order transfer functions with structure [1, 1, δ] representing fast and slow pathways, with characteristic time constants and steady-state gains, i.e. y(t)=bf1-afz-1ut-δ+bs1-asz-1u(t-δ)+ξt, where bf and bs are gains on the fast and slow pathways, respectively, and af and as are parameters characterising the time constants of the fast and slow pathways, respectively; af and as are roots of the denominator polynomial in the second-order transfer functions above (Eq. 2). This can be interpreted as two parallel linear storages.

In continuous time, a transfer function model with time delay τ has the following form: Y(s)=B(s)A(s)e-sτU(s)+E(s), where Y(s), U(s) and E(s) represent the Laplace transforms of the output, input and noise, respectively. A(s) and B(s) represent the denominator and numerator polynomials in the derivative operator s=ddt that define the relationship between the input and the output, and τ represents the time delay. second-order models were only accepted if they could be decomposed by partial fraction expansion into two parallel, first-order transfer functions, i.e. TPload=bfs+afe-sτR+bss+ase-sτR+E. This can be interpreted as two parallel stores, which are depleted at different rates, determined by the time constants (direct reciprocals of af and as) of the fast and slow components of the response, respectively. The terms bf and bs are parameters that determine the gain of the fast and slow components, respectively. The terms “fast” and “slow” are used here as qualitative terms, since they are not necessarily related to specific process mechanisms; for a second-order model (two stores), one store simply depletes at a slower rate than the other. Time constants are catchment specific; for example, for a first-order rainfall–runoff model which identifies just the dominant mode (one pathway), the time constant can vary from less than an hour (e.g. for a small, flashy catchment in Malaysian Borneo, Chappell et al., 2006) to more than 3 months (e.g. for a chalk stream in Berkshire, UK, Ockenden and Chappell, 2011).

This method of model identification requires high temporal resolution data that capture the dynamic response to the driving input; therefore, it cannot work if input data (in this case, rainfall) are missing, and does not perform well if too much output data (in this case, discharge or TP load) are missing or not showing a response. For the Newby Beck catchment, linear models were identified for short storm sequences up to 1 month, and were considered applicable to periods of similar conditions. These short-term models had a simple linear structure and very few parameters (five for a second-order model). As this paper is evaluating a methodology, successful modelling on different timescales can be used as validation of the approach. Models were not identified for short periods for Blackwater and Wylye, as the presence of a much slower pathway (with a time constant of the same order as the length of the identification period) did not allow model parameter estimates to be sufficiently constrained over such short periods.

For longer time series, when seasonal change and antecedent wetness are expected to have an impact on the response, linear models were improved by inclusion of the rainfall–runoff non-linearity (Beven, 2012) based on the storage state of the catchment, for which the discharge is used as a proxy, i.e. Re(t)=R(t)(Q(t-1))β, where Re(t) is the effective rainfall at time t, R is the observed rainfall, Q is the observed discharge, and β is a constant exponent that is optimized from the observed data at the same time as model identification. Using a simple non-linear function (with a single and optimised parameter) of recent discharge measurement as catchment wetness surrogate has been tested on catchments of different size and nature (e.g. Beven, 2012; Chappell et al., 1999; McIntyre and Marshall, 2010; Young, 2003; Young and Beven, 1994). A recent high flow will be highly correlated with high “overall” catchment wetness, and using the flow at time t-1, as in Eq. (6), still allows estimation of Re and Q at time t. The resulting effective inputs are rescaled in fitting the b parameters of the transfer function within the DBM calibration process. A transfer function model is not subject to a direct mass balance constraint, for example in flood forecasting applications where rainfall may be modelled against stage rather than discharge (e.g. Leedal et al., 2013). A simple antecedent precipitation index (API) was also tried initially, although this introduces additional parameterisation; it worked with reasonable success for Newby Beck but not for the other catchments, and therefore, as a consistent method was sought for all catchments, the API approach was not pursued in this case. For annual TP loads, the models (still with hourly timestep) were identified based on the data for hydrological years 2011–2012 and 2012–2013 for Newby Beck, but, because of missing output data, just for hydrological year 2012–2013 for the Blackwater and Wylye catchments. Models were validated on the data for all, or part, of the hydrological year 2013–2014.

Model fit was assessed according to model bias, to evaluate systematic over- or under-prediction of the model, and to Rt2 (also known as the Nash–Sutcliffe efficiency): Rt2=1-σ^2σy2, where σ^2=1N∑i=1Nyi-y^2;σy2=1N∑i=1Nyi-y‾2;y‾=1N∑i=1Nyi. Here, y^ is the model simulation, σ^2 is the mean squared error of the model residuals (only equal to the variance if the mean of the residuals is zero), and σy2 is the variance of the observations, yi. A balance of model fit and over-parameterisation was sought using the Young information criterion (YIC) and visual inspection of the model fit to the monitoring data. Model assessment criteria are defined in Sect. S2.

Time-series data from each catchment (Fig. 2) indicated large contrasts in the hydrological response of each study catchment, with Newby Beck (Eden) showing a very flashy response to rainfall (Fig. 2a). Although a fast response at certain times was also evident in the Blackwater (Wensum) catchment (Fig. 2c) and the Wylye (Avon) catchment (Fig. 2e), there was also a more pronounced seasonal response, particularly in the Wylye where a large groundwater component could be observed in the winter periods. This indicates the importance of both high-frequency data and a long-term record, to capture both fast and slower dynamics adequately. The errors resulting from sampling well below the catchment dynamics have been well documented elsewhere (e.g. Johnes, 2007; Jones et al., 2012; Lloyd et al., 2016b; Moatar et al., 2013). TP concentrations in all three study catchments revealed peaks that corresponded with runoff, with maximum values of 1.0, 0.9 and 1.5 mg L-1 in the Newby Beck, Blackwater and Wylye catchments, respectively. Newby Beck showed a very low background concentration of TP at low flow (minimum < 0.01 mg L-1), compared to 0.05–0.1 mg L-1 in the Blackwater, and around 0.12 mg L-1 in the Wylye. The relationships between streamflow and TP concentration are shown in Figs. S1–S3 in the Supplement, and the relationships between streamflow and TP load are shown in Figs. S4–S6. The presence of a measurable, background, non-rainfall-dependent concentration suggests an additional source of phosphorus to the recently applied agricultural sources. Such non-rainfall-dependent sources include legacy stores of agricultural P in the soil, both large and smaller point source discharges, such as sewage treatment works and domestic septic tanks (Zhang et al., 2014), and groundwater, specifically contributions from mineral sources in the Upper Greensand geology of the Hampshire Avon (Allen et al., 2014).

A summary of the observed total rainfall, runoff, mean concentration and TP load is given in Table 1 for the period 1 October 2012–30 September 2013 (the hydrological year with the most complete dataset). The lowest mean annual TP concentrations were observed in the Newby Beck catchment, but combined with the highest runoff this resulted in a high total annual TP load. Conversely, although mean annual TP concentration in the Blackwater was also higher than in Newby Beck, when combined with the lowest runoff, this resulted in the lowest total annual TP load. The rainfall–runoff ratio for Newby Beck (0.65) was much higher than for the Blackwater (0.31) or the Wylye (0.32), indicating a larger capacity for storage in the latter two catchments. Despite similarity in the rainfall–runoff ratio, total runoff in the Wylye was higher than the Blackwater because of the higher total rainfall.

Detailed analysis of the high-frequency data is not included here as it has already been published by several authors (e.g. Ockenden et al., 2016; Outram et al., 2014, including hysteresis analysis; Perks et al., 2015). Investigation of the relationships between TP concentration and streamflow indicated that, for all three catchments, the TP concentration was out of phase with the streamflow; distinct hysteresis loops (Figs. S1–S3), also observed by Outram et al. (2014), showed different TP concentrations on the rising stage of a storm hydrograph compared to the same stage on the falling hydrograph. This indicates that antecedent conditions and the storage state of the catchment are important in determining the response. In order to capture the effects of storage, dynamic models are required.

For short storm sequences up to about a month, when antecedent flows for events were rather similar, linear models were identified for the Newby Beck catchment. These were useful for infilling missing discharge or TP load data, or for highlighting and estimating uncertainties in discharge and TP load when extrapolation of the stage–discharge relationship was inappropriate. The model is only reliable for the conditions covered during the calibration period, but it may still be useful when there are known problems with a stage–discharge relationship (such as during extreme events). Indeed, the stage to discharge relationship is the weakest point of all the catchment models relying on stage measurements. Whilst it was still possible to identify linear models for short periods for the Blackwater and Wylye catchments, the parameter uncertainty for these models was large; the parameters cannot be well constrained when the (slow) time constant was of similar order to the period of identification. For this reason, linear models for short periods for the Blackwater and the Wylye were not considered useful.

Table 2 shows results from rainfall–runoff and rainfall–TP load models identified for Newby Beck for a series of contiguous storms in November 2015, immediately preceding Storm Desmond (5–6 December 2015), which caused catastrophic flooding in Cumbria and Lancashire, UK. During Storm Desmond, Honister Pass in Cumbria received the highest 24 h rainfall on record (341 mm) and Thirlmere received the highest 48 h rainfall on record (405 mm). The storm was remarkable for the duration of sustained rainfall. At Newby Beck, 156 mm of rainfall was recorded in 36 h. Although the monitoring equipment was recording during Storm Desmond, the peak flows during the storm were out of bank for around 31 h (compared to less than 3.5 h during more typical storms), with anecdotal evidence that the gauging point was significantly bypassed, so these out of bank flows were highly uncertain. This measurement uncertainty is shown by the shaded bands in Fig. 3 (discharge model) and Fig. 4 (TP load model), which span the observed (calculated from stage) discharge and TP load. This is more visible in the zoomed-in periods for discharge (Fig. 3b) and TP load (Fig. 4b). Concentrations were assumed to be reasonably accurate, but discharge was likely underestimated, therefore TP loads were consequently underestimated too. Storm Desmond was not included in the model identification period. Using the models from the November period to simulate flows (Fig. 3) and TP load during Storm Desmond (Fig. 4) suggests that both discharge and TP load were underestimated. Time series and histograms of the residuals are given in Fig. S7 for discharge and Fig. S8 for TP load. The zoomed-in period for the TP load model (Fig. 4b) suggests that whilst the transfer function model got the timing of the load peak and the decay approximately right, the model generally started to respond before the observed load responded.

Although there are uncertainties associated with whether it is valid to extend the models identified above to an extreme event such as Storm Desmond, we believe that this highlights the possible underestimation in discharge and TP load during Storm Desmond and that the models in Table 2 might provide more realistic estimations of the true values.

Longer-term models, based on 2 years of hourly data, were identified for each catchment. Model fits (Rt2) for rainfall–runoff models for the identification period (Table 3) were 0.71 for Newby Beck and 0.87 for Wylye, but only 0.37 for the Blackwater. Model bias was less than ±10 % for all three catchments. The runoff models were all linear transfer function models relating effective rainfall to discharge, where the exponent in the non-linear relationship between rainfall and effective rainfall (Eq. 6) was optimised at the same time as model parameter identification. The non-linearity, which reflects the effect of the antecedent soil moisture conditions in the catchments, was accounted for with the soil moisture surrogate expressed as a power function of discharge (Beven, 2012) with exponent β in Eq. (6), where a value of zero produces a linear response to rainfall and a higher value leads to an increasingly non-linear response. The β values identified for Newby Beck, Blackwater and Wylye were 0.37, 0.65 and 0.59, respectively, indicating the most non-linear response was in the Wensum (Blackwater) catchment, which also gave the lowest model efficiency values. The best identified model for rainfall–runoff in each catchment was a second-order model. In general, models higher than second order gave little improvement in model fit but a large deterioration in YIC, signifying over-parameterisation not warranted by the information in the monitoring data, whereas first-order models often gave a reasonable fit to the model peaks (and hence reasonable Rt2), but poor fit to recession periods.

The dynamic response characteristics of time constant and percentage on each flow pathway (for definitions see Table S4), determined after partial fraction decomposition, can be compared between the study catchments for both discrete and continuous-time models. The time constants are associated with the dominant pathways and indicate how quickly each impulse response (of water or TP mass) is depleted to 37 % (or fraction 1/e) of the peak exported. This is the standard definition of a time constant in a first-order linear time-invariant dynamic process, e.g. A(t)=A0exp⁡(-t/Tc), where Tc is the time constant. In reality there will be a continuum of runoff pathways with different time constants (Kirchner et al., 2000), but the information in the data indicates that this continuum can be simplified by representation as just two dominant pathways.

The marginal distributions of the time constants and proportion of flow or TP load (Table 3) were determined from 1000 to 10 000 Monte Carlo realisations using the covariance of the parameter estimates. The parameter uncertainties estimated within the DBM methodology were small, even for the response characteristics of the TP load models, which had higher uncertainty than rainfall–runoff models; TP load models had coefficients of variation of less than 3 % for fast time constants, less than 6 % for slow time constants and less than 2 % for proportions on pathways. For the rainfall–runoff models, the time constant for the fast pathway was 2.9 ± 0.1 h for Newby Beck, with 43 ± 0.5 % of the water taking this pathway; in the Wylye, the time constant for the fast pathway was 4.1 ± 0.2 h, but with only 8 ± 0.2 % of the water taking this route. This is consistent with the much higher baseflow index in the Hampshire Avon (0.93) than the Eden (0.39) (Table S1), which is clearly visible in the data (Fig. 1). For the Blackwater, 25 ± 0.6 % of the flow took the fast pathway, which is also consistent with the baseflow index in the Wensum (0.8) being between the Eden and Hampshire Avon. The fast time constant for the Blackwater catchment was much slower, at 14.8 ± 0.25 h; this may be related to the average slope of the catchment, which is much lower for the Blackwater catchment (less than 2 %) compared to 6–8 % for the Wylye and Newby Beck catchments. The slow time constant for Newby Beck was 147 ± 5 h, with 57 ± 0.5 % of flow taking this pathway; this compared with 441 ± 13 h (75 ± 0.6 % of flow) for the Blackwater and 395 ± 6 h (92 ± 0.2 % of flow ) for the Wylye.

For the rainfall–TP load models, at Newby Beck the best identified model was a first-order model relating the effective rainfall (from the runoff model, i.e. calculated one step at a time using the simulated discharge, Qsim) to the TP load (Table 3, Fig. 5). Although it was possible to identify a second-order model, this made virtually no difference to model fit, Rt2, whilst making YIC more negative (signifying over-parameterisation), and decomposition of the model revealed time constants for the two pathways that were both less than 8 h (cf. 147 h for the slow pathway for the rainfall–runoff model in Table 3). This indicates that in Newby Beck, all the TP load is transported through a quick flow pathway. This is consistent with most of the load being associated with P mobilised from diffuse agricultural sources, which is transferred by surface runoff or shallow sub-surface flow. This includes particulate P transported in surface runoff or drain flow (Heathwaite et al., 2006), subsurface movement of fine particles and colloids (Heathwaite et al., 2005), and displacement of fast subsurface soluble P sources. Young (2010) recommended a minimum data sampling rate of one-sixth of the time constant in order to avoid possible temporal aliasing effects. Littlewood and Croke (2013) illustrated the parameter inaccuracy and loss of data when observations were under-sampled for discrete time transfer functions, with inaccuracy decreasing and parameter estimates approaching stable values as the sampling interval decreased from 24 h (daily sampling) down to hourly sampling. The time constant for the first-order TP load model for Newby Beck was 1.6 ± 0.04 h. In this study, daily data would not capture the true dynamics of discharge and TP load, and that, ideally, for flashy catchments such as Newby Beck, a sampling interval shorter than hourly would be even more robust. However, for the other catchments in this study, the hourly data frequency was sufficient. The time constant for the TP load model (1.6 ± 0.04 h) was even faster than the fast time constant for the second-order (two pathway) rainfall–runoff model (2.9 ± 0.1 h), indicating that the TP mass impulse response was depleted at a faster rate than the water response, i.e. that the store was diluted as the storms progressed or that the sources must be readily connected and closer to the stream, since TP depends on transport velocities and we would normally expect velocities to be less than celerities under wet and surface runoff conditions. Those source areas would also be the most readily exhausted, so the effects would reinforce each other.

Expanded sections of Fig. 5 are shown for storms in May 2012 (Fig. 6a) and November 2012 (Fig 6b). Time series of residuals and residuals against observed values are given for the discharge model in Fig. S9 and for the TP load model in Fig. S10. Although Fig. 5 illustrates several storms where the model underestimated the peak TP load, the model matched the shape and peak of the May 2012 storm quite well. However, once again the model started to respond to the rainfall before the observations showed a response. Figure 6b shows an example of a storm in which the TP load was underestimated by the model. The model parameter uncertainty was considerably smaller than the measurement data uncertainty. The model did not always lie within the bands indicated by the measurement data uncertainty, whereas the total model prediction uncertainty (including the residual uncertainty) would span most of the observations, indicating that the simple structure of the model does not capture all the dynamics, and that there are other sources of uncertainty (such as rainfall input) which are not quantified.

For the Wylye, the best identified TP load model was a second-order model relating effective rainfall to TP load, with 42 ± 1 % on a fast pathway (TC = 6.1 ± 0.3 h) and 58 ± 1 % on a slower pathway (570 ± 54 h) (Table 3, Fig. 7). Compared to the runoff model, this showed a much greater percentage of the TP load on faster pathways such as surface runoff, shallow sub-surface flow or sub-surface drains. Nevertheless, there was still a significant proportion travelling on a slower pathway, which highlights the need for pollution mitigation efforts to include measures that take into account sub-surface and groundwater flows and, also, to recognise that surface runoff from farmland is not the only source of nutrients and sediment (Allen et al., 2014; Evans, 2012). These models cannot provide spatial information, but having identified that a slow pathway is so important, measures which prevent pollutants getting to the slow pathway in the first place, such as reductions at source, will be helpful. This may require further specific measurements, such as testing P in soils or identifying septic tanks in the catchment. With DBM models, this interpretation is made a posteriori, after the data assimilation, and is based on inferences from the objectively identified dominant modes of the system response.

Figure 8 shows expanded sections of the Wylye TP load model, including a large storm in which the load is underestimated (Fig. 8a) and two smaller storms where the model overestimated the loads (Fig. 8b). For the Wylye catchment, the measurement uncertainty was dominated by the uncertainty of the data from the TP sensor, rather than the uncertainty in the discharge (Lloyd et al., 2016b). However, some of the mismatch between model and observations here might also be attributable to uncertainty in rainfall input: in Fig. 8a there could be an underestimate in catchment rainfall not captured by the rain gauges; conversely, in Fig. 8b the rain gauges may have captured more than the catchment-average rainfall. Time series of residuals and residuals against observed values are given for the Wylye discharge model in Fig. S11 and for the TP load model in Fig. S12.

The TP load model used for the Blackwater was a linear model relating rainfall directly to TP load. The second-order TP model gave fast and slow time constants of 12.5 ± 0.6 and 376 ± 44 h, respectively (Table 3, Fig. 9). The time constants were similar in magnitude to, though both slightly shorter than, the time constants for the runoff model, suggesting a possible exhaustion effect where, as in Newby Beck, the TP mass store was diluted as the response progressed. For the Blackwater, as in the other study catchments, the proportion of TP load transferred on the fast pathway (54 ± 2 %) was considerably more than the proportion of water on the fast pathway (25 ± 0.6 %). Although seasonal non-linearity was still evident in the data from Blackwater, the rainfall–runoff models that included the non-linearity did not validate the data very well (Fig. S18), such that the two-stage TP models using the effective rainfall calculated one step at a time using the simulated discharge, Qsim, gave a worse fit to the data than a simple linear model. This may have been due to missing data in the discharge and TP time series, particularly over the storm peaks, or to inadequate representation of P inputs. An expanded section of Fig. 9, showing a series of storms in December 2012 (Fig. 10a) indicates the seasonal non-linearity of the response, which cannot be captured with a linear model, with a linear rainfall input. The first storm was considerably underestimated, but later storms were overestimated. This can usually be accounted for by using a non-linear effective rainfall input, which was unsuccessful in this case. A storm in May 2013 (Fig. 10b), when the land might have been drier than during the December storms, showed considerable overestimation of TP load by the linear model fitted to the December period. Time series of residuals and residuals against observed values are given for the Blackwater discharge model in Fig. S13 and for the Blackwater TP load model in Fig. S14.

The proportion of TP load exported on the fast pathway was considerably greater for all catchments than the corresponding proportion of water on the fast pathway, by a factor of approximately 2 for Newby Beck and Blackwater and approximately 5 for the Wylye. This suggests that on the fast water pathways, generally associated with shallower pathways such as shallow sub-surface flow, field drains and surface runoff, there is more release of TP than on deeper water pathways. This is consistent with soil profiles in agricultural areas, which generally show P concentrated on the surface and in the near-surface soil layers, with a decrease in P with depth (Heathwaite and Dils, 2000).

Validation of the TP model for Blackwater and Wylye was performed on a shorter period than for Newby Beck (half of the hydrological year 2013–2014) because of missing data (Table 3, Figs. S15–S18). The power law used to represent the rainfall–runoff non-linearity did not validate the data very well in the Blackwater catchment. Different representations of the rainfall–runoff linearity were also investigated, such as the Bedford Ouse Sub-Model (Chappell et al., 2006; Young, 2001; Young and Whitehead, 1977), in which the soil storage is related to an antecedent precipitation index. Although changes in the model non-linearity representation made minor differences to model fit, none of the model variants validated the data well for the Blackwater catchment. This suggests that there may be a different mechanism at work in the Blackwater catchment, in which a fast pathway only becomes active once the soil is fully saturated, or the groundwater level rises to a certain level (Outram et al., 2016). This could be due to the shallow slopes, which encourage infiltration rather than runoff. Alternatively, the response may be more dominated by point sources which are not as rainfall-driven, or sources such as sediment-laden runoff from impervious surfaces (roads or yards), which are rainfall-driven but do not behave in the same non-linear way as the runoff from soil.

In addition, the conditions experienced during the 2 years used for model identification may not be very similar to the validation period. From the data in Fig. 1c, the winter of 2011 and spring of 2012 showed much lower discharge than the same months in subsequent years. The groundwater recharge, which is shown as an increase in the baseflow in winter, was obvious for winter 2012–2013 and winter 2013–2014 for both the Blackwater (Fig. 2c) and the Wylye (Fig. 2e), but was not evident for either catchment for the winter of 2011–2012. Because of the slow time constants for these catchments, the dataset for model identification ideally needs to be longer than for the Newby Beck catchment, where the dynamics are much faster. This study suggests that the dataset used here was not long enough for the Blackwater catchment to capture an adequate range of conditions.

The benefits and limitations of the modelling method for TP load are summarised in Table 4. For catchments that exhibit rapidly changing dynamics, such as response to storm events, models calibrated with daily data will have large uncertainties associated with the parameters (and output) because the input data do not capture the high-frequency dynamics of processes such as P transfer. This study shows that simple transfer function models using data with sub-daily resolution can simulate the dynamics of TP load, with model fits at least as good as generally achieved with process-based models (Gassman et al., 2007; Moriasi et al., 2007) and with low parameter uncertainty. Full direct model comparisons are not currently possible, as the published results for process-based models used different catchments and data sets. It is still advisable to validate a fitted model using at least a split record test (Klemes, 1986). This highlights the importance of long and complete datasets with good time resolution for properly representing both flow and TP loads for such catchments. The high data demand of DBM models is noted in Table 4. Technology and monitoring methods are improving all the time so that high-frequency data are now more readily available (e.g. Jordan et al., 2007, 2005; Outram et al., 2014; Skeffington et al., 2015) This requirement for adequate datasets is often an obstacle in the use of the DBM modelling method, but as such datasets become more available, the method can be used to improve our understanding of catchments. We should embrace efforts to improve data coverage and ways to use it widely.

The models in Table 3 have been identified using a consistent method, to test how well this modelling method copes with the different characteristics of the three catchments. The method has been successfully applied to all the catchments, although less successfully for the Blackwater catchment. It is likely that the models could be improved if catchment-specific adjustments were made or used alongside other models in a hypothetico-inductive manner (Young, 2013). For instance, in the Blackwater catchment, the use of state-dependent parameters (Young, 1984) might be more successful to capture the rainfall–runoff non-linearity. This means that, rather than using the form of the non-linearity specified by Eq. (6), the parameters could be allowed to vary according to some other observed state. In addition, model fit might be improved by accounting for heteroscedasticity of residuals (shown in residual analysis, Figs. S9–S14), through transformation of data and residuals (e.g. Yang et al., 2007). Models for all catchments could be improved by having a longer dataset, to ensure, as far as possible, that environmental conditions during a future simulation period have already been experienced during the identification period.

The models showed a pattern of underestimation of high-level TP load events and, to a lesser extent, overestimation of lower level events (Figs. 10, 12 and 14). This was more apparent for TP load than for the discharge model (Figs. 9, 11 and 13), although in many cases this was within the limits of the uncertainty in the observed data. This suggests that, for the TP load model, the non-linearity may be rainfall, discharge or load-dependent to a greater extent than allowed for in the non-linearity of Eq. (6). This could be explored using state-dependent parameter estimation, on which the power law of Eq. (6) for the flow non-linearity was originally based (Young and Beven, 1994; Young, 1984). In addition, models with at least two terms in the numerator polynomial could provide more flexibility for a differencing effect, i.e. a consistent flushing effect with higher load occurring during the rising limb of the discharge peak. This mechanism is not represented in first-order models [1 1 del], as for Newby Beck, as it requires two terms of the numerator polynomial.

The use of process-based models is often justified on the basis that the inclusion of adequate process representations will lead to more robust estimation of the response to changing environmental conditions. This is the basis for arguing that process-based models are better suited for predicting the impacts of future change. However, they also involve a plethora of (often difficult to validate) assumptions in their model structures and parameters. In practice, parameters set during calibration are rarely changed to account for changes in the modelled processes under future conditions, although by calibrating models for conditions similar to the expected future conditions, it may be possible to incorporate non-stationary parameter values (Nijzink et al., 2016). This idea could be integrated into DBM models by choosing identification periods which are most likely to reflect the conditions of the simulation period or through the use of state-dependent parameters. Thus, whilst the data-based assumption of similar conditions may be questioned when limited periods have been used for identification, usually restricted by data availability, we argue that many of the factors contributing to catchment response will not have changed (e.g. catchment topography, soil type and geology) and that this assumption will in many circumstances be no more restrictive than the (different) assumptions made when using process-based models. Clearly, where the factors contributing to catchment response have obviously changed (such as if all septic tanks were upgraded or if farm budgeting reduced the additions of P), then simple transfer function models would not be able to predict the changes over time, whereas, in theory, process-based models might be able to account for such changes, albeit with much uncertainty (e.g. Dean et al., 2009; Yang et al., 2008). However, for rainfall-dominated responses, or responses to changes in rainfall patterns, simple transfer function models can provide valuable understanding of the dominant modes of a catchment, which, in turn, can be used to target management interventions.