A data mining approach to predictive vegetation mapping using probabilistic graphical models

This paper develops a novel method to model and predict the spatial distribution of vegetation types in Swaziland using physiographic and bioclimatic variables. The method uses a data mining approach implemented within a probabilistic graphical model to match two observed hierarchical levels of vegetation. The classification uses Bayesian networks (BN) and the parameterization is based on the expectation-maximization (EM) algorithm. The model is tested on a random sample of mapped vegetation types in Swaziland and allowed for the identification of the key environmental variables that are most important for capturing the vegetation spatial distribution. We show that while elevation and geology are the most important variables explaining the spatial distribution patterns of vegetation for both models, the influence of the climatic and other variables on the vegetation at the two levels differ. The overall distribution of the predicted vegetation classes was very similar to their distribution on the observed vegetation map. Overall the error rate was found to be 9.35% for a model of 16 vegetation classes and 4.9% for the one with 5 classes, indicating the excellent classification accuracy of the approach despite the complex landscape of the study area. Possible sources of error and some limitations are discussed and conclusions are drawn including suggestions for further investigation.