A supervised training and learning method for building identification in remotely sensed imaging

The paper investigates a novel approach for building identification in aerial images, that combines a classical segmentation algorithm, the region growing algorithm, a user guided training approach and a supervised learning solution based on support-vector machines. The user is guiding the training procedure by choosing points on the surface of objects of interest, e.g. buildings, as well as points over objects that are of no interest for the application, e.g. streets or vegetation. A local region growing algorithm is applied at each location chosen by the user. The system then prompts the user to label the type of object he/she selected. At the same time, a global region-growing algorithm is applied at uniformly spread seeds over the image and the resulting regions are combined. A series of features based on shape are then built for each region and a support-vector machine is trained to classify between objects of interest versus objects of no interest. The proposed solution obtains results in line in terms of recall and better in terms of precision than those reported in the remote sensing literature.