@InProceedings{KhoshelhamOude:2012:RoDiRe,
               author = "Khoshelham, Kourosh and Oude-Elberink, Sander",
                title = "Role of dimensionality reduction in segment-based classification 
                         of damaged building roofs in airborne laser scanning data",
            booktitle = "Proceedings...",
                 year = "2012",
               editor = "Feitosa, Raul Queiroz and Costa, Gilson Alexandre Ostwald Pedro da 
                         and Almeida, Cl{\'a}udia Maria de and Fonseca, Leila Maria Garcia 
                         and Kux, Hermann Johann Heinrich",
                pages = "372--377",
         organization = "International Conference on Geographic Object-Based Image 
                         Analysis, 4. (GEOBIA).",
            publisher = "Instituto Nacional de Pesquisas Espaciais (INPE)",
              address = "S{\~a}o Jos{\'e} dos Campos",
             keywords = "Disaster management, Pattern classification, Feature selection, 
                         Detection, Complexity, Lidar, Segmentation.",
             abstract = "We present a segment-based approach to detecting damaged building 
                         roofs in aerial laser scanning data. It consists of a segmentation 
                         step, where points are grouped into planar segments, a feature 
                         extraction step, and a classification step, where each segment is 
                         classified as damaged or intact. Such a segment-based approach 
                         faces two major challenges: first, extraction of features that are 
                         relevant to the target classes and can adequately distinguish 
                         between the intact and damaged segments is not straightforward. 
                         Second, the generation of reference segments for training and 
                         testing is difficult due the complexity of interpreting point 
                         clouds. To overcome these challenges the role of feature selection 
                         and dimensionality reduction in training a classifier using few 
                         training samples is investigated. We evaluate the performance of 
                         several classifiers with different sets of features in terms of 
                         classification accuracy. The results indicate the usefulness of 
                         dimensionality reduction methods in segment-based classification 
                         of aerial laser scanning data with few training samples. With 12 
                         features and 50 training segments a linear classifier outperforms 
                         more complex classifiers; however, dimensionality reduction 
                         methods result in larger improvements in the performance of 
                         complex classifiers.",
  conference-location = "Rio de Janeiro",
      conference-year = "May 7-9, 2012",
                 isbn = "978-85-17-00059-1",
             language = "en",
         organisation = "Instituto Nacional de Pesquisas Espaciais (INPE)",
                  ibi = "8JMKD3MGP8W/3BTG832",
                  url = "http://urlib.net/ibi/8JMKD3MGP8W/3BTG832",
           targetfile = "103.pdf",
                 type = "Urban Applications",
        urlaccessdate = "07 maio 2024"
}