Euro-limpacs Deliverables

ABSTRACT - DELIVERABLE 69

Report summarising the preliminary investigation of model uncertainty

This report describes;

1. a range of methods for assessing uncertainty in environmental models,

2. the key aspects of model structure and model−parameters that control the behaviour of four river−basin scale models used in work packages 4 and 6, and

3. preliminary estimates of uncertainty associated with the model predictions of flow, water chemistry and ecological changes.

As such, this work provides a basis of the uncertainty analysis that will be used in the later stages of the project to estimate uncertainty in the modelled forecasts of the response of freshwater ecosystems to expected changes in climate and other environmental drivers, such as land−use and atmospheric deposition. Specifically, the work done was as follows:

A statistical framework using Bayesian Modelling and Markov Chain Monte Carlo methods for model−calibration utilizing information in long−term monitoring data, including uncertainties, is presented. The statistical framework was linked to the hydrogeochemical model MAGIC used for predicting recovery from acidification in response to reduced acid deposition. Long term (29 years) monitoring data from Norway were used in the model calibration and evaluation. Uncertainties in model parameters were estimated and the model was calibrated to match observed soil chemistry and long−term runoff chemistry and runoff amount. Forecasts with uncertainties were presented for water quality and probability of fish survival for three different scenarios. Under the current legislation scenario, the probability distribution for having a sustainable trout population had a maximum density around 0.25 in year 2020, while the corresponding distribution for the scenario with no anthropogenic contribution peaked at 0.9. There were clear differences between the scenarios despite considerable overlapping probability distributions. The presentations of the water chemistry results as probabilities were thought useful to non−specialists; the results convey the likelihood of occurrence whereas simple error−bands around a water−chemistry time−series still require significant interpretation. Therefore it is recommended that the uncertainty in model forecasts (and hindcasts) be presented as probabilities of occurrence where appropriate. Monte Carlo runs coupled to a General Sensitivity Analysis (GSA) were used to identify the key parameters controlling the behaviour of INCA−Sed. Of the hydrological parameters, the proportions of saturation and Hortonian direct runoff delivered to the channel, and the time constant associated with soil through−flow were found to be most sensitive. Of the sediment parameters, the parameter associated with the release of in−situ fines and the transport capacity were the most sensitive. It is not surprising that these parameters were most significant in controlling the suspended sediment concentrations in channel reaches, as they control the principal source of sediment to the channel.

The performance of the NAM, N and P loss−models and MIKE11−TRANS was evaluated in the Gjern river−system using split−sample testing. All three approaches were reasonable for N but further work is required to improve predictions of P.

The sensitivity of the INCA−P model was analysed used Monte−Carlo techniques coupled with evaluation of model performance using the mean average error. The preliminary results suggest that the key parameters, in a groundwater catchment, are the initial flow of groundwater, the groundwater time−constant, the size of the initial in−organic P pool and the transfer of P between ?labile? and ?non−active? stores in the soil. Thus in this study there were four investigations of model uncertainty and performance testing, and in each a different technique was applied, in two examples to achieve the same aim, namely, a sensitivity−analysis of a model?s structure and parameters. It was apparent that there are numerous tools available with which to perform sensitivity and uncertainty analyses, three of which are reported on here. There are also at least two other well−known methods described in the contemporary literature: GLUE and SCEM. Given the diversity in approaches proposed in this report, it is recommended that a ?standard method? be considered for use in the Euro−limpacs project such that estimates of uncertainties between models might be more readily compared. In particular it is important to establish the similarities and differences of alternative methods of uncertainty estimation.

login to download Please login to download this report

« back to Deliverables