Building Detection From Pan-Sharpened Ikonos Imagery Through Support Vector Machines Classification

An approach is presented for detecting the buildings from high resolution pan-sharpened IKONOS imagery through binary Support Vector Machines (SVM) classification. In addition to original spectral bands, the bands nDSM (normalized Digital Surface Model), NDVI (Normalized Difference Vegetation Index), PC1, PC2, PC3, and PC4 (First, Second, Third, and Fourth Principal Components), are also included in the classification. The proposed classification procedure was carried out in three study areas selected in the Batikent district of Ankara, Turkey. The study areas show different residential and industrial characteristics. The first study area covers mainly the residential parts that include buildings with different shapes, sizes, dwelling types, and colored roofs. The second study area also represents the residential characteristics but contains buildings with more regular shapes. The third study area contains the industrial buildings with the gray tone roofs and the sizes of the buildings are larger. Also tested in the present study is the effect of the training sample size in the accuracy of the SVM classification. The results reveal that the overall accuracies were computed to be between 90% and 99%, while the kappa coefficients were found to be between 0.80 and 0.98. The inclusion of additional bands in the SVM classification had a considerable effect in the accuracy of building detection. Increasing the training size increased the accuracy, however, the increase was not more than 3%.Wo