Parameterizing Support Vector Machines for Land Cover Classification

The support vector machine is a group of relatively novel statistical learning algorithms that have not been extensively exploited in the remote sensing community. In previous studies they have been found to generally outperform some popular classifiers. Several recent studies found that training samples and input data dimensionalities can affect image classification accuracies by those popular classifiers and support vector machines alike. The current study extends beyond these recent research frameworks and into another important inquiry area addressing the impacts of internal parameterization on the performance of support vector machines for land-cover classification. A set of support vector machines with different combinations of kernel types, parameters, and error penalty are carefully constructed to classify a Landsat Thematic Mapper image into eight major land-cover categories using identical training data. The accuracy of each classified map is further evaluated using identical reference data. The results reveal that kernel types and error penalty can substantially affect the classification accuracy, and that a careful selection of parameter settings can help improve the performance of the support vector classification. These findings reported here can help establish a practical guidance on the use of support vector machines for land-cover classification from remote sensor data.