Mapping ecological systems with a random forest model: tradeoffs between errors and bias

Abstract

New methods for predictive vegetation mapping allow improved estimations of plant community composition across large regions. Random Forest (RF) models limit over-fitting problems of other methods, and are known for making accurate classification predictions from noisy, nonnormal data, but can be biased when plot samples are unbalanced. We developed two contrasting maps of forested ecological systems in the western Oregon Cascades ecoregion based on (a) RF and (b) RF with a bias adjustment. The methods had similar overall accuracy but different strengths and weaknesses. Both methods predicted dominant systems well. For systems with small sample sizes, accuracy was lower and differed more between methods. The bias adjustment process improved accuracy for minor systems with only minor impact on overall accuracy. The unadjusted RF model severely overestimated the area of abundant systems and underestimated minor classes. The adjustment process improved the areal estimates but did not completely eliminate the bias problem. Choice of methods and resulting maps should be based on objectives of the particular project.

Citation

Grossmann, E.; Ohmann, J.; Kagan, J.; May, H.; Gregory, M. 2010. Mapping ecological systems with a random forest model: tradeoffs between errors and bias. Gap Analysis Bulletin. 17: 16-22.