Final Report Summary - IMPROVE (Improving risk assessment in environmental decision making through robust uncertainty estimation)
Water and other environmental resources are threatened by human-induced changes such as pollution, climate change and over-exploitation. The sustainable management of these resources requires reliable information with the smallest possible uncertainty for robust decision-making. Improved methods for quantifying, reducing and communicating uncertainty are needed to minimise the socio-economic risks of decision making at local, regional and global scales. The overall purpose of this project was to improve risk assessment in environmental decision making through robust quantification, reduction and communication of uncertainties in data and model predictions. In particular the project aimed to develop methods to incorporate uncertainties in observational data into regionalised constraints of hydrological functioning in ungauged locations, and investigate how data and model uncertainties best can be communicated.
First, novel methods to estimate the uncertainty in observed hydrological data were developed. River discharge data are derived from water level (stage) measurements at most gauging stations using a rating curve, which is a model of the stage-discharge relation that is fitted to infrequent simultaneous measurements (gaugings) of stage and discharge (Figure 1). A new method for estimating rating-curve (and hence discharge) uncertainty was developed that accounted for random uncertainty in the discharge and stage measurements, and epistemic uncertainty (i.e. uncertainty related to lack of knowledge) about the true stage-discharge relationship (e.g. because of extrapolation, weed growth, and riverbed erosion). The method enabled the generation of multiple realisations of the rating curve parameters and, therefore, of the discharge data time series, using Monte Carlo sampling (Figure 1). It succeeded in representing uncertainty stemming from different causes of epistemic error and different types of rating curve equations across a wide range of catchments. The method is useful for hydrology practitioners to estimate discharge uncertainty and impacts on subsequent analyses.
We then investigated the effect of data uncertainty on uncertainty in hydrological signatures (i.e. indices quantifying different aspects of hydrological behaviour that are widely used in research and water management). First, a generally applicable method to estimate signature uncertainty based on Monte Carlo sampling was developed and applied to two catchments where detailed experimental data were available. This analysis included estimation of rainfall data uncertainty from subsampling of rainfall stations, and the methodology was applied to signatures of varying complexity. The results show that uncertainties are often large (i.e. typical intervals of ±10–40% relative uncertainty) and highly variable between signatures. There was greater uncertainty in signatures that use high-frequency responses, small data subsets, or subsets prone to measurement errors. There was lower uncertainty in signatures that use spatial or temporal averages. It was found that signatures could be designed to be robust to some uncertainty sources.
Discharge signature uncertainties were then estimated across a large dataset of UK catchments. The uncertainty varied with signature type, local measurement conditions and catchment behaviour. The highest uncertainties (median relative uncertainty ±30–40% across all catchments) occur for signatures measuring high and low flow magnitude and dynamics, which has important implications for drought and flood estimation and management. The lowest general uncertainties were found for groundwater-dominated catchments with dampened flow variability. Signature uncertainties of the magnitudes found through this research have the potential to change the conclusions of hydrological analyses and the interpretation of catchment functional behaviour.
The signature uncertainties were then propagated into predictions of signature values for catchments treated as ungauged (so called regionalisation), while accounting for uncertainty in the regionalisation procedure. Such predictions are an important information basis for water management since monitoring is not practically and economically feasible for all catchments. We found that there was a clear risk of over-conditioning the regionalisation inference if the gauged uncertainties were neglected; differences in signature values between catchments that are a result of data uncertainties could be wrongly attributed to differences in catchment behaviour. These research results have important implications for cross-catchment comparisons, catchment classification, and regional hydrological modelling.
Methods to communicate and visualise discharge data uncertainties were investigated. New ideas for communication of uncertainty related to hydrological systems, their model representations, and human-impacted change to the systems were discussed in a workshop and applied for an example water-management problem. Signature constraints were applied in hydrological modelling and hypothesis testing, where it was found that the information available about the data uncertainties is of central importance for reliable model evaluation, and that the type of signatures needed depend on the structure of the hydrological model. The developed methodologies allow for interesting future research in rainfall-runoff model regionalisation.
The project contributed with new methods and insights about observational uncertainties and their impact on the reliability of hydrological analyses in gauged and ungauged basins. The research results show that the impacts of observational uncertainty are large enough to have the potential to change the conclusions of hydrological analyses – with important consequences for water-resources management and decision-making, e.g. for floods, droughts, and nutrient load estimation. The new methods developed in this project have the potential to be used for a wide range of studies in the future – allowing better understanding of how uncertainties in fundamental hydrological data impact on conclusions about water resources of high societal importance.
First, novel methods to estimate the uncertainty in observed hydrological data were developed. River discharge data are derived from water level (stage) measurements at most gauging stations using a rating curve, which is a model of the stage-discharge relation that is fitted to infrequent simultaneous measurements (gaugings) of stage and discharge (Figure 1). A new method for estimating rating-curve (and hence discharge) uncertainty was developed that accounted for random uncertainty in the discharge and stage measurements, and epistemic uncertainty (i.e. uncertainty related to lack of knowledge) about the true stage-discharge relationship (e.g. because of extrapolation, weed growth, and riverbed erosion). The method enabled the generation of multiple realisations of the rating curve parameters and, therefore, of the discharge data time series, using Monte Carlo sampling (Figure 1). It succeeded in representing uncertainty stemming from different causes of epistemic error and different types of rating curve equations across a wide range of catchments. The method is useful for hydrology practitioners to estimate discharge uncertainty and impacts on subsequent analyses.
We then investigated the effect of data uncertainty on uncertainty in hydrological signatures (i.e. indices quantifying different aspects of hydrological behaviour that are widely used in research and water management). First, a generally applicable method to estimate signature uncertainty based on Monte Carlo sampling was developed and applied to two catchments where detailed experimental data were available. This analysis included estimation of rainfall data uncertainty from subsampling of rainfall stations, and the methodology was applied to signatures of varying complexity. The results show that uncertainties are often large (i.e. typical intervals of ±10–40% relative uncertainty) and highly variable between signatures. There was greater uncertainty in signatures that use high-frequency responses, small data subsets, or subsets prone to measurement errors. There was lower uncertainty in signatures that use spatial or temporal averages. It was found that signatures could be designed to be robust to some uncertainty sources.
Discharge signature uncertainties were then estimated across a large dataset of UK catchments. The uncertainty varied with signature type, local measurement conditions and catchment behaviour. The highest uncertainties (median relative uncertainty ±30–40% across all catchments) occur for signatures measuring high and low flow magnitude and dynamics, which has important implications for drought and flood estimation and management. The lowest general uncertainties were found for groundwater-dominated catchments with dampened flow variability. Signature uncertainties of the magnitudes found through this research have the potential to change the conclusions of hydrological analyses and the interpretation of catchment functional behaviour.
The signature uncertainties were then propagated into predictions of signature values for catchments treated as ungauged (so called regionalisation), while accounting for uncertainty in the regionalisation procedure. Such predictions are an important information basis for water management since monitoring is not practically and economically feasible for all catchments. We found that there was a clear risk of over-conditioning the regionalisation inference if the gauged uncertainties were neglected; differences in signature values between catchments that are a result of data uncertainties could be wrongly attributed to differences in catchment behaviour. These research results have important implications for cross-catchment comparisons, catchment classification, and regional hydrological modelling.
Methods to communicate and visualise discharge data uncertainties were investigated. New ideas for communication of uncertainty related to hydrological systems, their model representations, and human-impacted change to the systems were discussed in a workshop and applied for an example water-management problem. Signature constraints were applied in hydrological modelling and hypothesis testing, where it was found that the information available about the data uncertainties is of central importance for reliable model evaluation, and that the type of signatures needed depend on the structure of the hydrological model. The developed methodologies allow for interesting future research in rainfall-runoff model regionalisation.
The project contributed with new methods and insights about observational uncertainties and their impact on the reliability of hydrological analyses in gauged and ungauged basins. The research results show that the impacts of observational uncertainty are large enough to have the potential to change the conclusions of hydrological analyses – with important consequences for water-resources management and decision-making, e.g. for floods, droughts, and nutrient load estimation. The new methods developed in this project have the potential to be used for a wide range of studies in the future – allowing better understanding of how uncertainties in fundamental hydrological data impact on conclusions about water resources of high societal importance.
