Landscape Applications of Machine Learning: Comparing Random Forests and Logistic Regression in Multi-Scale Optimized Predictive Modeling of American Marten Occurrence in Northern Idaho, USA

Abstract

The American marten is a species that is dependent on old conifer forest at middle to high elevations and is highly sensitive to habitat loss and fragmentation. Our goal was to compoare logistic regression and random forest in multi-scale optimized predictive model of occurrence of the American marten (Martes americana) in northern Idaho USA. There have been relatively few formal comparisons of the performance of multi-scale modeling between logistic regression and random forest, but those that have been conducted have found that random forest out-performs logistic regression. There was substantial similarity in the qualitative interpretation of the logistic regression and the random forest model and both found that occurrence was strongly predicted by a unimodal function of elevation, a non-linear function of canopy cover, a non-linear function of patch density, and the extent of the landscape in large conifer forest. Visual inspection of the predicted occurrence probability maps shows that random forest produces predictions that are more discriminatory, with higher range of predicted probability and higher spatial heterogeneity than logistic regression. The logistic regression model has an AUC of 0.701, while the random forest model had an AUC of 0.981, indicating very high predictive ability, and a much stronger ability to predict presences and absences in the training dataset than the logistic regression model. Expressed as a percentage, the random forest model had 28% higher performance, leading to much better prediction of habitat suitability, better inferences about habitat variables influencing marten occurrence, improved identification of scale dependency, and ultimately, therefore, better guidance to conservation and management.