Object- and Image Endmember-based Riparian Forest Classification of Narrow-Band UAS Image Data: A Case Study of the River Gail and River Drau, Austria

Studies that directly compare classification accuracies of object-based image analysis (GEOBIA) and endmember-based algorithms for the exploitation of very-high-spatial-resolution (VHR) images (e.g., unmanned aircraft systems (UAS) images) are quite limited. We employ an endmember-extraction algorithm in conjunction with an endmember-mapping method, and we separately utilize a multiresolution segmentation/object-based classification algorithm. We then classify riparian forest and other land covers and compare the classification accuracies obtained from the application of these respective classifiers to narrow-band, VHR UAS images acquired over two river reaches (of the River Gail and River Drau, respectively) in Austria. We determine the effect of pixel size on classification accuracy and assess performances associated with multiple image-acquisition dates. Our results indicate markedly higher classification accuracies for the GEOBIA approach, relative to those of the endmember-based method, where the former generally entails overall accuracies in excess of 85%. Poor endmember-mapping classification accuracies are most likely a function of: the very small pixel sizes associated with the UAS images; the large number of information classes; and the relatively small number of (albeit narrow) bands available for analysis.