A comparison of pixel-based and object-oriented approaches to VHR imagery for mapping saltmarsh plants

a r t i c l e i n f o Object-oriented classification (OOC) has shown many significant advantages over other methods for classification of urban or forest ecosystems. However, it remains unclear if this technology could exhibit these advantages on mapping monospecific plant stands in herbaceous plant dominated ecosystems (e.g. saltmarshes). In this study, we compared the effectiveness of OOC and pixel-based classification (PBC) methods for mapping plants in a saltmarsh ecosystem. QuickBird was selected for very high resolution (VHR) imagery. Eleven models defined by classification types, feature spaces, classifiers, and hierarchical approaches with multi-scale segmentation were built for comparison. The results showed that the QuickBird imagery efficiently discriminated saltmarsh monospecific vegetation stands and that OOC performed better than PBC in terms of accuracy. We also found that the improvement of OOC was primarily due to employing membership functions and a hierarchical approach with multi-scale segmentation. Although texture and shape features have been deemed as two major advantages of OOC, enhanced performance was not observed in this study. The results of this study demonstrated that OOC would be superior to PBC for classifying herbaceous plant species in terms of accuracy. To improve the classification accuracy, greater concern should be given to exploration of the relationships between features of both objects and classes and to combining information from different object scales, while shape and texture features can be a minor consideration due to their intricately high spatial variability. Remote sensing offers a practical and economical means to mapping plant distribution over large areas. In recent decades, multi-spectral data such as TM and SPOT have been widely applied in wetlands to identify general vegetation classes or discriminate broad vegetation communi-However, images from TM or SPOT satellite instruments are insufficient for discriminating vegetation at the species level in herbaceous plant dominated environments (e.g., saltmarshes) due to a lack of high spatial resolution. Specifically, the diameters of plant patches in these habitats may be lower than those covered by the maximum average resolution of SPOT of 10 m and TM of 30 m, resulting in the formation of mixed pixels containing several plant species (Adam et al., 2010; He et al., 2010). Although the linear spectral mixture analysis (LSMA) (e.g., Shimabukuro and Smith, 1991) model has been proposed to enhance coarse image classification, this method can only produce the proportions of multiple materials of each pixel (or objects) rather than the actual distributions …