Classification of floodplain vegetation by data fusion of spectral (CASI) and LiDAR data

To safeguard the goals of flood protection and nature development, a river manager requires detailed and up-to-date information on vegetation structures in floodplains. In this study, remote-sensing data on the vegetation of a seminatural floodplain along the river Waal in the Netherlands were gathered by means of a Compact Airborne Spectrographic Imager ( CASI; spectral information) and LiDAR ( structural information). These data were used to classify the floodplain vegetation into eight and five different vegetation classes, respectively. The main objective was to fuse the CASI and LiDAR-derived datasets on a pixel level and to compare the classification results of the fused dataset with those of the non-fused datasets. The performance of the classification results was evaluated against vegetation data recorded in the field. The LiDAR data alone provided insufficient information for accurate classification. The overall accuracy amounted to 41% in the five-class set. Using CASI data only, the overall accuracy was 74% ( five-class set). The combination produced the best results, raising the overall accuracy to 81% ( five-class set). It is concluded that fusion of CASI and LiDAR data can improve the classification of floodplain vegetation, especially for those vegetation classes which are important to predict hydraulic roughness, i.e. bush and forest. A novel measure, the balance index, is introduced to assess the accuracy of error matrices describing an ordered sequence of classes such as vegetation structure classes that range from bare soil to forest.