Object-oriented land cover classification of lidar-derived surfaces

Light detection and ranging (lidar) provides high-resolution vertical and horizontal spatial data and has become an important technology for generating digital elevation models (DEMs) and digital surface models (DSMs). The latest terrestrial lidar sensors record intensity and echo information for each pulse in addition to range. In this study, lidar height and intensity data were used to classify land cover using an object-oriented approach. The study area was selected based on the variety of land cover types present and consists of urban, mixed forest, and wetland-estuary coastal environments. Surfaces constructed from the lidar points included DSM, DEM, intensity, multiple echos, and normalized height. These surfaces were segmented and classified using object rule based classification. Ten classes were extracted from the lidar data, including saturated and non-saturated intertidal sediments, saturated or stressed and lush ground cover vegetation, low and tall deciduous and coniferous trees, roads and bare soil, bright-roofed structures, dark-roofed structures, and water. The accuracy of the classification was assessed using independent ground reference polygons interpreted from colour orthophotographs and intensity images. The average accuracy of the 10 classes was 94%, but improved to 98% when the classification results were aggregated into seven classes. The results indicate that accurate land cover maps can be generated from a single lidar survey using the derived surfaces, and that image object segmentation and rule-based classification techniques allow the exploitation of spectral and spatial attributes of the lidar data.