Urban land cover mapping using medium spatial resolution satellite imageries: effectiveness of Decision Tree Classifier

The study is inserted in the framework of information extraction from satellite imageries for supporting rapid mapping activities, where information need to be extracted quickly and the elimination, also if partially, of manual digitalization procedures, can be considered a great breakthrough. The main aim of this study was therefore to develop algorithms for the extraction of urban layer by means of medium spatial resolution Landsat data processing; Decision Tree classifier was investigated as classification techniques, thus it allows to extract rules that can be later applied to different scenes. In particular, the aim was to evaluate which steps to perform in order to obtain a good classification procedure, mainly focusing on processing that can be applied to images and on training set features. The training set was evaluated on the basis of the number of classes to use for its creation, together with the temporal extension of the training set and input attributes, while images were submitted to different kind of radiometric pre and post-processing. The aim was the evaluation of the best variables to set for the creation of the training set, to be used for the classifier generation. Above-mentioned variables were compared and results evaluated on the basis of reached accuracies. Data used for the validation were derived from the Digital Regional Technical Map.