Transformation and texture based features in TerraSAR-X data classification for environmental monitoring

Feasibility of Random Forest and Support Vector Machine classifiers is tested for the discrimination of 7 types of vegetation near lake Poosjärvi in Western Finland. Four sets of features grouped as basic, textural, ICA or PCA based, and rotational features are applied. The results indicate that the Random Forest classification scheme outperforms the Support Vector Machine classifier. For both classifiers the textural features improve the performance significantly when added to the basic feature set while the ICA, PCA or rotational features have little effect. The best total classification accuracy of 87.5 % was obtained when all the considered feature sets were combined and fed to the Random Forest classifier.