A Land Cover Classification Method for Antarctica Using Support Vector Machine and Decision Tree

Global land cover data are fundamental for applications, especially ecological environmental assessment and climate change research. Currently available global land cover data products show some deficiencies in data accuracy and spatial and temporal resolution. So we discuss fast automatic classification methods for the study area in Antarctica. A classification method based on a Support vector machine (SVM) and a decision tree (DT) model is proposed. We compare the land cover classification using four common kernel functions for a SVM. The experiment indicates that the SVM classification method using a radial basis function (RBF) leads to the optimal accuracy and running time. In view of existing phenomenon that surface features in shadow areas are easily confused, classification is further improved by using a DT model, at last a majority analysis of the above classification result removes small polygon artifacts to form the final land cover data product. The overall accuracy is 95.82%, higher than the SVM alone and the maximum likelihood method. Land cover classification in Antarctica can be conducted more reliably through our proposed classification method.