Fusion of Airborne Hyperspectral and LiDAR Remote Sensing Data to Study the Thermal Characteristics of Urban Environments - Computational Approaches for Urban Environments - page 270

Fig. 11.2 The three-stage approach for surface material mapping

microclimate modeling. However, some target classes were also included because
the spectral and spatial resolution of the input data allowed for their classification
without introducing too many uncertainties in the final surface material map. For
the purpose of material mapping, a hybrid, three-stage classification approach was
employed (Fig. 11.2 ).
In the first stage, a series of segmentation algorithms was applied to the input
data (Berger et al. 2013 ). The goal of the segmentation step was to obtain larger
image objects for homogeneous regions, such as patches of grass or parking lots,
and smaller image objects for heterogeneous regions, such as densely built-up
areas. After the initial segmentation, a rule-based classification of the created image
objects was performed (Benz et al. 2004 ;TrimbleLtd. 2013 ) to generate a building
mask. The mask allowed the exclusion of spectrally similar ground classes (e.g.,
asphalt roads) from the later classification of roofing materials (Herold et al. 2004 ;
Herold and Roberts 2006 ;Herold 2007 ). Image segments were first divided into
elevated and non-elevated objects using the LiDAR nDSM. An object height of
2 m served as the separation threshold. The height value was chosen to enable
the differentiation between small but elevated objects such as allotment garden
cottages and pseudo-elevated objects such as vehicles (Ma 2005 ;Yuetal. 2010 ).
Subsequently, buildings were separated from all other elevated image objects.
Among the features used to derive the building mask were image brightness, the
NDVI, and the slope of the nDSM (Berger et al. 2013 ).
In the second stage, 15 samples of each target class were selected for supervised
data classification. The selection was based on spectral profile analyses, literature
comparisons, and visual inspection of the CASI and LiDAR data. In addition, the
training samples provided by the IEEE GRSS (Image Analysis and Data Fusion
Technical Committee 2013 ) were taken into account. On the basis of these samples,
two object-based support vector machines (SVMs) were established (Trimble Ltd.
2013 ). While the first SVM was trained to attribute one of four roofing materials
to each object in the building mask, the second SVM was set up to extract
the remaining target classes. Both SVMs were parameterized with a radial basis
function and took full advantage of all available or previously generated features of
the CASI and LiDAR data.