Object based analysis and decision tree for the classification of submetrical spatial resolution images captured by uav

Remotely sensed images acquired with submetrical spatial resolution, such as those captured by Unmanned Aerial Vehicle (UAVs), usually present a high spectral variability. An object based approach allows the generation of attributes, increasing the dimensionality of the original dataset. Selection functions for relevant features and booster, available in the C5.0 algorithm, and the objects based analysis facilitate the classification of such datasets. This study aimed at: (i) to evaluate object classification approach in relation to the parameters of attributes selection (winnow), booster (trial) and Minimum Sample Size (MSS), (ii) to determine the most predictive features, and (iii) to compare the classification performance of both Decision Tree and Support Vectors Machine. A region growing method was used in order to segment the image. C5.0 algorithm was used in the classification procedure. The values of the trial (10) and MMS (5) parameters resulted in accuracies higher than 0.8. The kappa value was greater that the SVM, with these parameters. A significant decrease in the data dimensionality up to 30% was observed when the winnow parameter was habilitated. The two most important features in the breakdown of classes were ‘ratio between the blue and green bands’ and ‘average of the elevation values’.

Keywords:
data mining; land cover mapping; Unmanned Aerial Vehicle (UAV).