Land Cover Classification at Continental to Global Scales using Kernel Machines

As the skin layer of the earth's surface, land cover plays important roles in many land surface processes. Land cover data is therefore required for studies of such processes. In particular, global change studies require land cover data at the global scale. Many global land cover products have been developed at moderate or coarse spatial resolutions using satellite imagery. Through these products, more scientific requirements have been realized for the designs of the classification stage. Continental to global scale mapping involves the characterization of land cover classes across long distances, which means that our classes are extrapolated from limited training data. This requires the classifier to have good generalization power across geographical variations. It has been noted in recent years that, the Support Vector Machines (SVM) classifier has better generalization power than Decision Trees, which has been the popular classifier used in global land cover classification. Support Vector Machines belong to a big family of classifiers known as the kernel machines. Another key difference between SVM and Decision Tree is that SVM integrates a design called 'regularization'. It was designed for classes with some overlapping in the feature space. However, it turns out to also provide us with some level of error tolerance towards training errors. This is a desirable feature for continental to global scale mapping. Current global land cover products derived from satellite imagery are rarely the results of direct classification of raw data. Instead, the classification is performed on a series of spectral 'metrics' derived from raw satellite data. Using these metrics helps to synchronize the vegetation phenology across northern and southern hemispheres, and also greatly reduces the dimensions and volume of data inputs. However, doing so also could over-simplify the information contents in the raw satellite data. Another member of the kernel machines classifiers, known as the Large Linear Classifier, has been designed by mathematicians specifically to handle data with very large dimensions. It shares some vital designs of SVM. With class generalization power, error tolerance, and high-dimensional data efficiency, the kernel machines are expected to improve land cover classifications. In this study, we will use MODIS data to evaluate these kernel machines as compared with conventional decision tree classifiers for producing continental to global scale land cover classifications, first in North America and then the whole world.