Mathieu Marmion

Statistical consensus methods for improving predictive geomorphology maps

A variety of predictive models is currently used to map the spatial distribution of earth surface processes and landforms. In this study, we tested statistical consensus methods in order to improve the predictive accuracy of geomorphological models. The distributions of 12 geomorphological formations were recorded at a resolution of 25ha in a sub-arctic landscape in northern Finland. Nine environmental variables were used to predict probabilities of occurrence of the formations using eight state-of-the-art modelling techniques. The probability values of the models were combined using four different consensus methods. The accuracy of the models was calculated using spatially independent test data by the area under the curve (AUC) of a receiver-operating characteristic (ROC) plot. The mean AUC values of the geomorphological models varied between 0.711 and 0.755 based on single-model techniques, whereas the corresponding values based on consensus methods ranged from 0.752 to 0.782. The weighted average consensus method had the highest predictive performance of all methods. It improved the accuracy of 11 predictions out of 12. The results of this study suggest that the consensus methods have clear advantages over single-model predictions. The simplicity of the consensus methods makes it straightforward to implement them in predictive modelling studies in geomorphology.

Assessing spatial uncertainty in predictive geomorphological mapping: A multi-modelling approach

Maps of earth surface processes and the potential distribution of landforms make an important contribution to theoretical and applied geomorphology. Because decision making often depends on information based on spatial models, there is a great need to develop methodology to evaluate the spatial uncertainty resulting from those models. In this study we developed a new method to produce maps of the uncertainty of predictions provided by ten state-of-the-art modelling techniques for sorted (SP) and non-sorted (NSP) patterned ground in subarctic Finland at a 1.0-ha resolution. Six uncertainty classes represent the modelling agreement between the different modelling techniques. The resulting uncertainty maps reflect the reliability of the estimates for the studied periglacial landforms in the modelled area. Our results showed a significant negative correlation between the degree of uncertainty and the accuracy of the modelling techniques. On average, when all ten models agreed, the mean area under the curve (AUC) values were 0.904 (NSP) and 0.896 (SP), these values decreased to 0.416 (NSP) and 0.518 (SP), respectively, when only five models agreed. Mapping of the uncertainty of predictions in geomorphology can help scientists to improve the reliability of their data and modelling results. The predictive maps can be interpreted simultaneously with the uncertainty information, improving understanding of the potential pitfalls of the modelling.