Using ensemble techniques to capture the accuracy and sensitivity of ecosystem service models

If the United Nations sustainable development goals (SDGs) are to be achieved, it is vital to understand the interactions between people and nature. A significant aspect of these interactions can be classed as ‘nature’s contributions to people’ (termed ecosystem services; ES). However, the empirical ES data needed to quantify these relationships are sparse in all parts of the World. Using recent advances in data availability from remote sensing, models are increasingly able to provide credible information where empirical data are lacking. Specifically, ES models produce maps of estimated ES (typically based on land cover and other driving variables) and so can provide the understanding of the spatial distribution and heterogeneity ES required to aid planning and optimisation of land use decisions. However, most ES modelling applications rely on a single model for each ES and few applications explicitly validate ES models against independent datasets. As a consequence, the uncertainties associated with each application of ES models (and the datasets that underpin) them remain largely unknown. This is a particular issue as the results of local scale validation are likely not to be transferable to new locations or to the regional and national scales at which ES model outputs are most widely used

EnsemblES seeks to address these issues by: 1) investigating ES model input sensitivity, varying initial conditions at the start of model simulations; 2) combining the outputs of multiple ES models (from multiple initial conditions) into ‘ensembles’ of models using a variety of techniques including when data on individual model performance is vs is not available; and 3) validating these model ensembles against independent data, highlighting a) the accuracy of ES ensembles, and b) whether coefficients of variation of the ensemble is a good predictor of ensemble uncertainty. 

The EnsemblES project will, for the UK context, identify whether ecosystem service (ES) model ensembles provide more accurate predictions than individual models. We will identify which method of combining models into an ensemble is most applicable for ES science, highlighting the sensitivity of models (and ensembles of models) to input data. Finally, we will evaluate the accuracy out these methods in data-deficient areas. Overall, we aim to highlight that although individual models can give good predictions, without validation against data or other models, this may impose a rigidity that might have serious negative consequences if the estimate deviates significantly from truth and so model ensembles should be used to support ES decisions. 

Principle Investigator Simon Willcock introduces the project in this short video
Project contacts:
Dr Simon Willcock: HomepageTwitter 
Prof James Bullock: HomepageTwitter 
Prof Laurence Jones: HomepageTwitter 
Project outputs:
1. Ensembles of ecosystem service models can improve accuracy and indicate uncertainty in Science of The Total Environment
Volume 747, 10 December 2020, 141006 https://www.sciencedirect.com/science/article/pii/S0048969720345356 the associated project dataset available under the terms of the Open Government Licence: https://catalogue.ceh.ac.uk/documents/11689000-f791-4fdb-8e12-08a7d87ad75f
Link to the project page on the UKRI Gateway to Research https://gtr.ukri.org/projects?ref=NE%2FT00391X%2F1