In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
Object
Completeness
(%)
Correctness
(%)
Quality
(%)
Buildings
98
72
60
Trees
99
83
80
Roads
78
72
62
Table 1: Accuracy Assessment
Figure 10: Extracted Objects & Orthophoto
Figure 11: 3D Model (County Sligo)
5. DISCUSSION
The classification results, indicated in Table 1, are the result of
automated processes, depending on the choice of appropriate
parameters and thresholds.
Building extraction is the first step in the classification process
and is therefore important for the extraction of further objects.
A completeness value of 98% implies that the adopted strategy
has been successful in the identification of these objects.
However correctness and quality values are significantly less
than completeness due to the influence of FP values. FP values
in buildings arise from large trucks or industrial installations
incorrectly identified as buildings. However, FN values (i.e.
missed buildings) mostly occurred for small buildings less than
50 sq. m.
In a subsequent manual step, these small buildings were
individually identified and included in the final building layer.
The identification of all buildings in the area in this way
allowed the extracted vegetation data to be improved, which
later helped in the extraction of roads. Commercial or
residential buildings, having glass roofs or green colour were
missed in the NDVI layer but they existed in the NDSM and
were added to the building layer manually. Many small sheds
were identified in the backyards of houses which are not part of
the buildings, which significantly reduced the correctness
value.
Vegetation was extracted by subtracting the building layer
from the NDSM. Very small buildings which appeared in the
vegetation layer were identified and manually added to the
building layer. Continuing research is targeted at reducing the
dependence on such manual steps.
Multiple reflections, size and compactness were used to
separate single trees from groves. However, the LiDAR sensor
can efficiently differentiate between multiple reflections only
where their height differences are significant.
Roads appear to be the most difficult objects to extract. They
are part of the DTM and have spectral reflectance, which varies
a lot in a single image. Setting a NDVI threshold helps identify
the areas where there is vegetation or not. Reflections from
barren land or walking trails in the fields also have very low
NDVI values. Roads which are not covered by building
shadows or trees are detected successfully. Road markings of
different colours also affect the extraction process. Roads
connecting houses to the road are of different materials and
need to be classified separately.
6. CONCLUSION
The accuracy of the generated orthophoto is critical for any
classification technique using LiDAR and aerial images. Due to
the nature of the push broom sensor and the configuration of
the test flight (no overlap along strip and 15% overlap between
strips) there is no possibility to combat limitations in the Red,
and NIR orthoimages. Occluded areas and ghosting of building
roofs (in the across flight direction) cannot be corrected
adequately and the building roof structure is completely
damaged in the areas close to strip edges. This is a major
disadvantage in the identification and modelling of building
roof structures. Ground control points, where available, should
be used for the verification of the registration of the LiDAR
point cloud and aerial images. In this approach we relied
completely on orientation from GPS\INS data but for future
research ground control points will be acquired and the
accuracy of the image registration will be measured.