Image Processing Reference
In-Depth Information
urban mapping, the building elevations can be removed in the last return signal using
different acquisition and processing techniques, i.e., minimum filters, large footprint
LIDAR data or existing ground elevation models. Then the difference between the
first and the last response elevations normalizes large-scale topographic variations
and emphasizes the three-dimensional surface structure of the urban environment
from buildings and vegetation. The LIDAR elevation difference can be used as a
pseudo spectral band in image classifications.
Another important consideration for reducing spectral confusion among urban
land cover types is spatial resolution. In fact, spatial and spectral resolutions are
strongly related since a distinct spectral signal can only be acquired if the spatial
resolution is sufficiently fine enough to represent the land cover object in “pure”
pixels. Figure 4.7 presents classification results to highlight spatial-spectral resolu-
tion effects, and the contribution of LIDAR information on the urban land cover
mapping process. In Fig. 4.7 , the producers accuracy measures the percentage of test
areas of a specific land cover type sampled in the real world that were classified
right. The users accuracy describes the percentage of all areas classified as a specific
class, and actually belong to this class in the sampled test data (Jensen 2000 ).
AVIRIS (14 most suitable bands) data shows significant spectral improvements as
indicated in the previous sections. However, including the LIDAR elevation difference
in the image classification provides an additional important level of information.
Especially for mapping buildings/roofs, the combination IKONOS/LIDAR resulted in
better classifications than using AVIRIS data. For most classes, the accuracy variations
for different sensor configurations are larger than the changes in spatial resolutions.
The producers accuracy of green vegetation is high for all sensor configurations and
spatial resolutions. The unique spectral signal of vegetation
is well represented in all sensor signals and, in terms of the
producers accuracy, allows for very accurate classification
results. The green vegetation users accuracy on the other
hand shows a tremendous decrease in accuracy for all sen-
sor types especially from 4 to 10 m spatial resolution, i.e.,
vegetation gets increasingly over-mapped at coarser spatial
resolutions. Pixels adjacent to green vegetation areas
increasingly merge with non-vegetation land cover types and
form mixed pixels. The strong spectral vegetation signal
leads to increasing amounts of vegetation being classified in the image. This trend is
evident for all sensor configurations and reflects the general limitations of lower
spatial resolution data in mapping urban land cover. For coarser spatial resolution,
the use of spectral mixture analysis helps to map urban land cover on the sub-pixel
level (Rashed et al. 2001 ).
In contrast to the other classes, bare soil classification accuracies show improve-
ments for lower spatial resolutions. Bare soil usually represents areas with larger
spatial extents that do not require high spatial resolutions for their accurate mapping.
The users accuracy also indicates the importance of the detailed spectral information
for accurate separation of bare soil from other land cover types. The signal from
IKONOS, and IKONOS and LIDAR elevation difference, is quite limited in this
the quality of
urban land cover
mapping strongly
depends on the
spatial and spec-
tral characteristics
of the remote
sensing data
 
Search WWH ::




Custom Search