2.1. Catchment conceptualization

In terms of the inputs and outputs that will form the basis of a machine learning model, we consider a crude representation of an inland catchment, one we assume is not affected by tidal inflows, with water entering the catchment through precipitation and exiting through evaporation, transpiration, and through the streamflow outflow, our target variable. We think of the internal processes as ones that move water toward one of the outflow processes or into internal storage.

More formally, the sum total of water that leaves a catchment via river discharge, that enters or leaves through climatic, or other, processes, and any changes to water stored within a catchment must be equal to zero, as per the integral form of continuity in Equation (2.1)) (Yilmaz et al., Reference Yilmaz, Gupta and Wagener2008), where the cumulative precipitation over a time period with length $ T $ in the catchment is given by $ {\int}_{t=0}^T{\Psi}_t dt $ , the water leaving the catchment through evapotranspiration is $ {\int}_{t=0}^T Edt $ , the target streamflow exiting through the river is $ \therefore {\int}_{t=0}^T{Y}_t $ , and the change in water storage is $ \Delta S $ .

(2.1) $$ {\displaystyle \begin{array}{l}\hskip6em 0=\int_{t=0}^T{\Psi}_t dt-\int_{t=0}^T{Y}_t dt-\int_{t=0}^T Edt-\Delta S\\ {}\therefore \int_{t=0}^T{Y}_t dt=\int_{t=0}^T{\Psi}_t dt-\int_{t=0}^T Edt-\Delta S\end{array}} $$

We further simplify the hydrological system by considering the meteorological inputs and outputs as acting at a single point, rather than being spatially distributed. This single point we take is the center of the catchment, or, more concretely, the centroid of the catchment, being the geometric center of a catchment’s two-dimensional surface, equivalent to its center of mass for a congruent shape of uniform weight density. Assuming that a catchment is a shape in two dimensions with area, $ A $ , its centroid, with coordinates $ \left(\overline{x},\overline{y}\right) $ , is given by taking first moments of area in both axes, with respect to some frame of reference, divided by the shape’s area, as in Equation (2.2)):

(2.2) $$ \overline{x}=\frac{\int_Ax\; dA}{\int_A dA};\hskip1em \overline{y}=\frac{\int_Ay\; dA}{\int_A dA} $$

where, for the purposes of this calculation, we will consider the catchment as a surface that exists in a flat, two-dimensional plane, effectively removing elevation and reducing the dimensionality of its boundary, $ B $ , such that $ B\in {\mathrm{\mathbb{R}}}^2 $ rather than $ B\in {\mathrm{\mathbb{R}}}^3 $ and treating that boundary as the continuous edge of an irregular shape with uniform density.

As we focus on catchment-specific models, we therefore do not use catchment descriptors as inputs. The predicted flow dynamics are dependent in a nonlinear way on the contemporaneous catchment descriptors, rather than stationary catchment descriptors, and the precipitation response function is not constant; rather, it is learned as a function of the antecedent conditions, captured through soil moisture or its proxies. Furthermore, many catchment descriptors within the United Kingdom are relatively stable; for example, the total amount of agricultural land, representing 72.8% of the UK’s total land cover, only varied by 1.3% in a 30-year period from 1990 to 2020 (Department for Environment, Food, and Rural Affairs, 2024; Department for Levelling Up, Housing and Communities, 2023).

2.2. Artificial neural networks

2.2.1. Overview

The multilayer perceptron (MLP) is composed of multiple, interconnected neuron units, each combining inputs linearly before an activation function is applied; depending on the choice for that activation function, nonlinearity can be introduced. A single neuron along the structure for an arbitrary MLP is shown in Figure 1, where the output, $ y $ , from a single neuron with respect to its inputs, $ \mathbf{x} $ , is a general linear combination of those inputs, with weights $ \mathbf{w} $ and bias $ b $ , prior to the activation function, $ \alpha $ , being applied.

Figure 1. Graphical representation of a single neuron unit and an arbitrary neural network with $ l $ layers.

If we let $ \Phi $ represent the set of all parameters to be learned, then we can use the difference between a prediction made with the network, $ {y}^{\prime } $ , and the target value, $ y $ , through a cost function, $ J\left(\Phi \right) $ , such as the root-mean-squared error (RMSE) form in Equation 2.3.

(2.3) $$ J\left(\Phi \right)={\left(\frac{1}{2n}\sum \limits_{i=1}^n{\left({y}_i^{\prime }-{y}_i\right)}^2\right)}^{\frac{1}{2}} $$

To minimize the value of the cost function, with respect to the parameters $ \Phi $ , we typically use gradient descent to update each parameter, $ {\phi}_j $ , according to the sign and magnitude of the gradient, controlled by the learning rate hyperparameter $ \eta $ , as per Equation 2.4.

(2.4) $$ {\phi}_j:= {\phi}_j-\eta \cdot \frac{\partial J\left(\Phi \right)}{\partial {\phi}_j} $$

The Backpropagation algorithm (Rumelhart et al., Reference Rumelhart, Hinton and Williams1986) enables the application of gradient descent in networks by propagating the derivatives backward between layers; for the parameters at the $ {l}^{th} $ layer of a network with $ L $ layers, the update gradient for that layer is chained back through all previous layers and activation functions, $ {\alpha}_L,\dots, {\alpha}_l $ , as in Equation 2.5.

(2.5) $$ \frac{\partial J\left(\Phi \right)}{\partial {\Phi}_l}:= \frac{\partial J\left(\Phi \right)}{\partial {\alpha}_L}\cdot \frac{\partial \left({\alpha}_L\right)}{\partial {\Phi}_{L-1}}\cdot \dots \cdot \frac{\partial \left({\alpha}_l\right)}{\partial {\Phi}_l} $$

2.3. Error metrics

To assess the performance of the equation, we employ the following metrics: RMSE, expressed in Equation 2.6; mean percent relative error (MPRE), expressed in Equation 2.7; relative bias (RB), expressed in Equation 2.8; Nash-Sutcliffe efficiency (NSE), expressed in Equation 2.9; and Kling-Gupta efficiency (KGE), expressed in Equation 2.10. The RMSE and MPRE both give an indication of the absolute magnitude of the absolute error and, accordingly, have a range of $ \left[0,\infty \right) $ _, where $ 0 $ indicates a perfect score. RB, on the other hand, identifies whether or not a model is on average under- or overpredicting and has a range of $ \left(-\infty, \infty \right] $ . NSE and KGE, through normalization, enable the model’s accuracy to be compared across catchments more easily and are commonly used in hydrology to do so (Gupta et al., Reference Gupta, Kling, Yilmaz and Martinez2009; Nash and Sutcliffe, Reference Nash and Sutcliffe1970), each with a range of $ \left(-\infty, 1\right] $ . While both metrics provide a sense of how the model is performing compared to the mean, where an NSE of 0 and a KGE of −0.41 would indicate performance equivalent to the mean flow benchmark (Wouter et al., Reference Wouter, Knoben, Freer and Woods2019), KGE is intended to capture the discrepancy between the variability and timing of the compared datasets. While obtaining a score of 1 $ 1 $ for the NSE and KGE would be ideal, obtaining scores of NSE $ \ge 0.5 $ (Moriasi et al., Reference Moriasi, Arnold, Van Liew, Bingner, Harmel and Veith2007) and KGE $ \ge 0.5 $ (Rogelis et al., Reference Rogelis, Werner, Obregón and Wright2016) are thresholds of suitable model performance.

(2.6)

(2.7)

(2.8)

(2.9)

(2.10) $$ KGE=1-\sqrt{{\left(r-1\right)}^2+{\left(\frac{{\overline{y}}^{\prime }}{\overline{y}}-1\right)}^2+{\left(\frac{\sigma_{y^{\prime }}}{\sigma_y}-1\right)}^2} $$

All of these metrics are expressed in terms of the $ n $ pairwise observations, $ {y}_i $ , and model predictions, , with the KGE also relying on the use of the linear correlation between observations and simulations, $ r $ , and the standard deviation of predictions, $ {\sigma}_{y^{\prime }} $ , and observations, $ {\sigma}_y $ .

2.4. Sensitivity and comparative analysis

To make a comparison of the impact that the soil moisture and antecedent proxy variables have, we utilize an additional pair of methodologies. The first is the correlation between variables, specifically the Pearson Correlation Coefficient, $ {r}_{xy} $ , in Equation 2.11, which we use to determine the information similarity between the soil moisture variables and the proxies.

(2.11) $$ {r}_{x_1{x}_2}=\frac{\sum_i\left({x}_i-\overline{x}\right)\left({y}_i-\overline{y}\right)}{\sqrt{\sum_i{\left({x}_i-\overline{x}\right)}^2{\sum}_i{\left({y}_i-\overline{y}\right)}^2}} $$

The second methodology is on the sensitivity of a network to the input variables, for which various methodologies exist. Commonly employed methods for doing so include, but are not limited to: input perturbation analysis, weight analysis, and gradient analysis (Cao et al., Reference Cao, Alkayem, Pan, Novák and Rosa2016; Cao and Qiao, Reference Cao and Qiao2008; Gevrey et al., Reference Gevrey, Dimopoulos and Lek2003; Pizarroso et al., Reference Pizarroso, Portela and Muñoz2022). The method that we employ here is the input perturbation algorithm, where for each input variable, $ {x}_i $ , we add some increment, $ {\delta}_i $ , that we can vary and then measure the change in output, where $ {\delta}_i $ is a proportion of the input value.

4.1. Variable comparison

We first provide analysis of the correlation between the antecedent proxies with the soil moisture variables, in accordance with the methods as described in Section 2, to highlight the information value of these variables and the potential utility of the antecedent proxies. The correlation between soil moisture and the antecedent proxies is shown in Table 1, where $ {SM}_{Lx} $ is the soil moisture at the level $ x $ and $ {\mu}_{v:d} $ is the antecedent proxy for a variable $ v $ , either precipitation or temperature, taken over a number of days $ d $ .

Table 1. Pearson correlation coefficients between soil moisture-level variables and antecedent proxies, averaged over the three study catchments

While the soil moisture-level variables exhibit positive correlation among themselves, the correlation between the deepest level, 4, and the shallowest level, 1, is weak. This suggests that there is significant information gain in utilizing all soil moisture levels simultaneously. (This does not mean that the trained neural network will respond to all variables equally: that quantification is shown in the subsequent sensitivity analysis.)

Key to the premise of this work is the strength of correlation between the antecedent proxies and the soil moisture levels. The results are intuitive: shorter-term antecedent proxies have high correlation (positive for precipitation and negative for temperature) with the shallow soil moisture levels; longer-term antecedent proxies exhibit high correlation with the deeper soil moisture levels. It is clear that different information is contained within the antecedent proxies across the timescales. Because we can reason physically that both shallow and deep soil moisture states are important for flow prediction (though the weighting of importance will be catchment dependent) due to their relationship to surface runoff and groundwater recharge, we are justified in selecting antecedent proxy timescales that correlate strongly with both deep and shallow soil moisture states.

4.2. Model results

Our results for the three catchments across all metrics are shown in Table 2, respectively, including results for the models using a meteorological input sequence length of 28 days, 7 days, 7 days with soil moisture, and 7 days with proxy variables. The performance measured by RMSE and MPRE indicates that the highest performing models are those that use the 7-day meteorological input with the inclusion of either the antecedent proxies or soil moisture variables. Increasing the number of days in the meteorological input space continued to increase performance, although we found that the gains in performance started to taper off beyond 28 days and came at the expense of numerical stability during training. From the RB, all models had a tendency to underpredict, though the evidence for which underpredicts the least was inconclusive.

Table 2. MLP model performance in terms of Nash-Sutcliffe efficiency for meteorological input sequence length of 28 days (28), 7 days (7), 7 days with soil moisture (7 + SM), and 7 days with proxy variables (7 + P) (bold typeface indicates highest performance)

Through examination of the NSE and KGE, however, the overall performance for the Findhorn at Shenachie is clearly lower than that for the other catchments. This was expected due to the smaller size of the catchment potentially leading to a more rapid flow response. Compared to the Bedford Ouse at Roxton, the Findhorn at Shenachie is smaller yet the mean flow rate is higher. The inference is that the response from the catchment requires a finer temporal scale than the daily scale used here. Regardless, we can still see the same pattern emerging that the short variable record in conjunction with either soil moisture or the antecedent proxies results in a model with a high degree of skill. Furthermore, all models exceed our benchmarks and the higher KGE relative to the NSE indicates that the models are predicting peaks and are representing the variability of the flow patterns well.

We present a subset of the results, those for the Severn at Haw Bridge: with the predictions against observations in Figure 3; exemplar streamflow time series observations and predictions for the year 2012 (from the test set) in Figure 4; and the percentage relative error for different percentile bands in Figure 5. The figures for the other catchments are shown in the Supplementary Materials.

Figure 3. Predictions against observations for the Severn at Haw Bridge from the test set of predictions generated using different feature sets as inputs to the MLP model.

Figure 4. Comparative streamflow time series for the Severn at Haw Bridge in the year, 2012, with both predictions and observations using different feature sets as inputs to the MLP model.

Figure 5. Violin plots of percentage relative error for each of the 10 $ {}^{th} $ percentile bands of flow magnitude (with the upper bound marked on the scale and the lower bound being the preceding upper bound to the left) between observations and predictions for the Severn at Haw Bridge using different feature sets as inputs to the MLP model.

Across all three catchments, baseflow and peak flow events are less well captured in the 28-day and 7-day models; whereas this is noticeable for winter peaks in the larger two catchments, all model implementations struggle with the summer peaks for the Findhorn at Shenachie. Using the soil moisture and proxy variable models result in a significant improvement in the representation of baseflow and a more modest improvement in the representation of peak flow events. Model results, while not perfect, are approaching a level of performance that would be considered robust. The set of catchments tested here is, obviously, small and further work may be required to refine the approach further. Having said that, with this framework in place, an alternative departure point from this work might be to move into the multiple catchment setting to prove that this is a suitable climatic framework for constructing a generalizing model; in that context, the catchment descriptors would come into play and, perhaps with more data, enable the learning of small catchment response versus larger catchment response.

4.3. Network sensitivity

Using the input perturbation algorithm, we assess the sensitivity of the two separate networks trained on either the soil moisture variable set or the antecedent proxy variable set. Sensitivity has been calculated with perturbations in either direction, to account for the varying response to lower temperatures and higher precipitation when compared with lower precipitation and higher temperatures; arguably, utilization of the gradient method of sensitivity analysis might avoid this issue and will be explored in further work. A subset of results averaged across the three study catchments is presented in Table 3. Model sensitivity to the soil moisture variables and the antecedent proxy variables is of a similar order of magnitude; if we consider the aggregate sensitivity to these variable sets, then it is roughly equivalent over these two variable sets for the positive perturbation but higher for the antecedent variable set for the negative perturbation. Essentially, the response of the network to these variables appears to be similar. In terms of whether or not all the proxies or soil moisture levels are required to force the model, due to the varying information represented and the relative sensitivities, then it would be hard to argue that any one variable was not required in either the soil moisture to antecedent proxy variable sets, though soil moisture Level 1 might perhaps be well represented by and surplus to soil moisture Level 2 and short-term precipitation.

Table 3. Average model sensitivity to individual soil moisture and antecedent proxy inputs using positive and negative perturbations for the two different model setups

When examining the response due to the daily meteorological variables, it is worth noting that the Findhorn at Shenachie, as a smaller, less permeable and therefore flashier catchment, is far more sensitive to the most recent 24 hours worth of rainfall when compared to the larger catchment that feeds the Severn at Haw Bridge. Overall, the closer to the flow event, the more impactful the perturbation. The network sensitivity drops off substantially after 5 days (backward from the event), as shown in Figure 6, and sensitivity to daily meteorological forcing is superseded by sensitivity to the soil moisture levels or antecedent proxies after this point. Thus, the antecedent proxies are capable of replacing longer meteorological records and the soil moisture levels, likely encapsulating much of the information encoded within the aforementioned variables.

Figure 6. Average network sensitivity to daily meteorological variables under a positive perturbation (left) and negative perturbation (right), with logarithmic scales on the y axis.

4.4. Study limitations

We note two areas in particular that were beyond the scope of this paper but are worth future exploration. The first being that, although we have quantified the error across percentile ranges, we have not developed the model to allow for prediction and reconstruction of the flow volume into base and peak flows. Though the output data we used were average daily flow, empirical methods exist for the reconstruction of base and peak flows that can be parameterized or trained on a per catchment basis (Fathzadeh et al., Reference Fathzadeh, Jaydari and Taghizadeh-Mehrjardi2017; Jimeno-Sáez et al., Reference Jimeno-Sáez, Senent-Aparicio, Pérez-Sánchez, Pulido-Velazquez and Cecilia2017). The second area is on the quantification of uncertainty, in terms of both the aleatoric and epistemic uncertainty. Inadequate quantification of uncertainty can hinder the utilization of artificial intelligence approaches in decision and policy-making processes (Borrego et al., Reference Borrego, Monteiro, Ferreira, Miranda, Costa, Carvalho and Lopes2008; Cabaneros and Hughes, Reference Cabaneros and Hughes2022) and provision of model uncertainty therefore aids adoption.