Random forests as a tool for ecohydrological distribution modelling

An important issue in ecohydrological research is distribution modelling, aiming at the prediction of species or vegetation type occurrence on the basis of empirical relations with hydrological or hydrogeochemical habitat conditions. In this study, two statistical techniques are evaluated: (i) the widely used multiple logistic regression technique in the generalized linear modelling framework, and (ii) a recently developed machine learning technique called ‘random forests’. The latter is an ensemble learning technique that generates many classification trees and aggregates the individual results. The two different techniques are used to develop distribution models to predict the vegetation type occurrence of 11 groundwater-dependent vegetation types in Belgian lowland valley ecosystems based on spatially distributed measurements of environmental conditions. The spatially distributed data set under investigation consists of 1705 grid cells covering an area of 47.32ha. After model construction and calibration, both models are applied to independent test data sets using two-fold cross-validation and resulting probabilities of occurrence are used to predict vegetation type distributions within the study area. Predicted vegetation types are compared with observations, and the McNemar test indicates an overall better performance of the random forest model at the 0.001 significance level. Comparison of the modelling results for each individual vegetation type separately by means of the F-measure, which combines precision and recall, also reveals better predictions by the random forest model. Inspection of the probabilities of occurrence of the different vegetation types for each grid cell demonstrates that correct predictions in central areas of homogeneous vegetation sites are based on high probabilities, whereas the confidence decreases towards the margins of these areas. Threshold-independent evaluation of the model accuracy by means of the area under the receiver operating characteristic (ROC) curves confirms good performances of both models, but with higher values for the random forest model. Therefore, the incorporation of the random forest technique in distribution models has the ability to lead to better model performances.