Object-based random forest classification for mapping floodplain vegetation structure from nation-wide CIR and LiDAR datasets

Very high resolution aerial images and LiDAR (AHN2) datasets with a national coverage provide opportunities to produce vegetation maps automatically. As such the entire area of the river floodplains in the Netherlands may be mapped with high accuracy and regular updates, capturing the dynamic state of the vegetation. In this study, these fused datasets are used to map the vegetation of 936 ha of the floodplain on the north-side of the river Nederrijn near Wageningen into ten vegetation structure classes. The method follows object-based image analysis principles. Objects are defined in segmentation and subsequently labeled using the ensemble-tree classifier random forest. The mapping scale is controlled by selecting segmentation parameters from quantified discrepancies between reference polygons and segmented objects. Effects on the mapping scale of different reference polygons and different segmentation data is investigated. The results show that it is important to be able to select the right segmentation parameters to control the mapping scale. A discrepancy measure with reference polygons is a suitable method to do this objectively. The use of random forest classification on the objects resulted in an estimated classification accuracy of 86% on the basis of the built-in cross-validation estimate of random forest. Variable importance measures of random forest showed that the AHN2 lidar dataset is a valuable addition to the spectral information contained in the aerial images in the classification.