Image Processing Reference
In-Depth Information
Table 4.2 Overall classiication accuracies for mapping 26 urban land cover classes in three different
spectral sensor resolutions in 4 m spatial resolution (simulated IKONOS and Landsat TM, and
AVIRIS data represented by 14 most suitable bands shown in Fig. 5, Herold et al. 2003 )
accuracy (%)
coefficient (%)
Area weighted
accuracy (%)
Built classes
accuracy (%)
IKONOS (4 bands)
Landsat TM (6 bands)
AVIRIS (14 bands)
multispectral sensor configurations (e.g., IKONOS and Landsat TM) can be
simulated from hyperspectral data. The results for simulated IKONOS and Landsat
TM data, and AVIRIS data using 14 most suitable bands classifying 26 urban
land cover classes (Maximum Likelihood classification) are shown in Table 4.2 .
The classification results for simulated IKONOS and Landsat TM versus
AVIRIS (selected 14 bands) show the improvement in map accuracy with increasing
spectral resolution. The area-weighted accuracies are higher since major parts of
the study area are covered with vegetation; these areas
map with little error. The differences between IKONOS
and AVIRIS accuracy ranges from ~12% for the mean
overall accuracy and Kappa to ~15% for the area weighted
accuracy and nearly 30% for only the built categories. The
improvements for built-up class mapping clearly confirm
the limitations of IKONOS in detailed separation of urban
land cover types. Landsat TM data provided intermediate accuracies with much
better performance than IKONOS in particular for the built land cover classes; the
improvements in classifying these cover types was more than 16% compared to
IKONOS. This especially highlights the importance of the SWIR region. The four
bands in the VIS/NIR region of IKONOS and Landsat TM have similar spectral
coverage so the two additional SWIR bands in Landsat TM are the main reason for
the difference in classification accuracy (Table 4.2 ).
As expected, AVIRIS with the 14 most suitable bands shows the best mapping
performance. While IKONOS and Landsat TM are broadband multispectral sys-
tems, the narrow AVIRIS bands can resolve small-scale absorption features and
the increasing number of bands better separate more cover types. In general, the
spectral sensor characteristics of IKONOS and Landsat TM were designed for
mapping a variety of land surfaces, especially for acquisition of natural and
quasi-natural environments. Different spectral sensor configurations are antici-
pated to resolve the unique spectral properties and complexity of urban environ-
ments. However, the AVIRIS land cover classification of 26 different urban land
cover classes illustrated general limitations in mapping the urban environment
even using hyperspectral optical remote sensing data. These limitations reflect
the similar spectral characteristics of certain land cover types indicated in inter-
pretation of spectral signatures and separability analysis. Considering the high
degree of within-class variability due to roof geometry, condition, and age, their
classification accuracy was low, reaching only 66.6% for the built up categories
(Table 4.2 , Herold et al. 2003 ).
landcover map-
pings accuracy
increases with
increasing spectral
Search WWH ::

Custom Search