Gradient modeling of conifer species using random forests

• Published in: Landscape ecology 2009-05, Vol.24 (5), p.673-683
• Main Authors: Evans, Jeffrey S, Cushman, Samuel A
• Format: Article
• Language: English
• Subjects: Animal, plant and microbial ecology; Applied ecology; Biological and medical sciences; Biomedical and Life Sciences; Coniferous trees; Ecology; Environmental Management; Forestry; Fundamental and applied biological sciences. Psychology; General aspects; General aspects. Techniques; Landscape; Landscape Ecology; Landscape/Regional and Urban Planning; Life Sciences; Methods and techniques (sampling, tagging, trapping, modelling...); Nature Conservation; Research Article; Species composition; Sustainable Development; Vegetation patterns
• ISSN: 0921-2973; 1572-9761
• DOI: 10.1007/s10980-009-9341-0

Landscape ecology often adopts a patch mosaic model of ecological patterns. However, many ecological attributes are inherently continuous and classification of species composition into vegetation communities and discrete patches provides an overly simplistic view of the landscape. If one adopts a niche-based, individualistic concept of biotic communities then it may often be more appropriate to represent vegetation patterns as continuous measures of site suitability or probability of occupancy, rather than the traditional abstraction into categorical community types represented in a mosaic of discrete patches. The goal of this paper is to demonstrate the high effectiveness of species-level, pixel scale prediction of species occupancy as a continuous landscape variable, as an alternative to traditional classified community type vegetation maps. We use a Random Forests ensemble learning approach to predict site-level probability of occurrence for four conifer species based on climatic, topographic and spectral predictor variables across a 3,883 km² landscape in northern Idaho, USA. Our method uses a new permutated sample-downscaling approach to equalize sample sizes in the presence and absence classes, a model selection method to optimize parsimony, and independent validation using prediction to 10% bootstrap data withhold. The models exhibited very high accuracy, with AUC and kappa values over 0.86 and 0.95, respectively, for all four species. The spatial predictions produced by the models will be of great use to managers and scientists, as they provide vastly more accurate spatial depiction of vegetation structure across this landscape than has previously been provided by traditional categorical classified community type maps.