Improving the automated classification of aerial imagery

Autores UPV
CONGRESO Improving the automated classification of aerial imagery


One of the main practical issues when applying data mining techniques for land cover classification of aerial imagery is the large amount of attributes used for describing the data. In addition to the increase of the learning time, another drawback is that this represents a source of noise. But, are all these attributes necessary to learn a good classifier? In this paper, we carried out an experimental analysis trying to improve the land cover classification of aerial imagery. We focused on urban and peri-urban areas, used an object-oriented image segmentation approach in order to better face the classification problem, and tested three different spatial resolutions (0.5, 1, and 2 m/pixel). Then, we analysed the attributes that were more relevant for the classification and how they improved the performance of the model. However, once the model has been learnt, a new question arises: can we reuse it in a different place? In a second experiment, we investigated whether the models that have been generated in a particular location can be used to classify a new geographical area with similar land cover types. The results show that the model reusing performs quite well for some type of classes.