Variable importance for sustaining macrophyte presence via random forests: data imputation and model settings

Abstract

Data sets plagued with missing data and performance-affecting model parameters represent recurrent issues within the field of data mining. Via random forests, the influence of data reduction, outlier and correlated variable removal and missing data imputation technique on the performance of habitat suitability models for three macrophytes (Lemna minor, Spirodela polyrhiza and Nuphar lutea) was assessed. Higher performances (Cohen’s kappa values around 0.2–0.3) were obtained for a high degree of data reduction, without outlier or correlated variable removal and with imputation of the median value. Moreover, the influence of model parameter settings on the performance of random forest trained on this data set was investigated along a range of individual trees (ntree), while the number of variables to be considered (mtry), was fixed at two. Altering the number of individual trees did not have a uniform effect on model performance, but clearly changed the required computation time. Combining both criteria provided an ntree value of 100, with the overall effect of ntree on performance being relatively limited. Temperature, pH and conductivity remained as variables and showed to affect the likelihood of L. minor, S. polyrhiza and N. lutea being present. Generally, high likelihood values were obtained when temperature is high (>20 °C), conductivity is intermediately low (50–200 mS m−1) or pH is intermediate (6.9–8), thereby also highlighting that a multivariate management approach for supporting macrophyte presence remains recommended. Yet, as our conclusions are only based on a single freshwater data set, they should be further tested for other data sets.