Regional land cover classification from MODIS time-series and geographical data using support vetor machine

The study investigated the performance of support vector machine (SVM) classifier for regional land cover mapping. First, 8 input features derived from MODIS time series and DEM data were selected by Jeffreys-Matusita distance. Then, all the features were analyzed to generate land cover map of Sanjiang Plain in China, using SVM algorithm. Finally, we evaluated the impact of sample size and its distribution on classification accuracy. The train and test ratio of 8:2 was proved to be a better choice for improving land cover classification. The distribution of samples influenced classification results, with a standard deviation of 0.81 to overall accuracy and 0.01 to Kappa coefficient. The overall accuracy of resultant classification map was 96.45 with Kappa coefficient of 95.8%. The good performance indicated great potentials of SVM algorithm for regional land cover mapping.