You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Fusion of sensor data, knowledge sources and algorithms for extraction and classification of topographic objects
Baltsavias, Emmanuel P.

International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
5.2. Data processing
The DEMs have been processed with commercial software
packages. The SPOT DEMs are part of the DEM of the full
scene, which has been generated by the Leica Helava DPW 770
digital photogrammetric workstation, which uses cross
correlation for point matching. The correlation coefficient of
each DEM is not output directly, but is hidden behind a figure
of merit (FOM). This FOM is scaled in a range from 0 to 100
and is directly proportional to the correlation coefficient (Leica
Helava, 1997). Height values below a FOM threshold of 33 are
generally considered unreliable. Although the images were
taken in a four-day interval, the SPOT data had to be enhanced
by filtering. Further details of the SPOT pre-processing and the
accuracy of the full scene DEM are given in (Patias, 1998).
The InSAR DEM has been generated with the PCI IFSAR
package. IFSAR offers both the LS and tree algorithms for
phase unwrapping, but as the tree algorithm failed for test site 2
and delivered the same RMS in site 1, LS has been applied to
both regions, to keep the results comparable. Due to lack of
ground control points (GCPs) in both areas, 5 GCPs had to be
derived from the reference DEM for geocoding and baseline
fitting. Baseline fitting has been performed by a technique
proposed by Werner (1992). The amount of GCPs reduced
remarkably the systematic error in the InSAR DEMs in
comparison to earlier results with less GCPs (Honikel, 1998).
The SPOT and ERS DEMs have been compared to ground truth
by bilinear interpolation of each height value in the reference
DEM. All referred errors are absolute errors, unless something
different is stated. Detailed investigations of the influence of the
crop height on the measurements have not been carried out.
Still, the fact that the signed average is larger in site 2 and
negative indicates a bias between the tree canopies, measured in
the DEMs, and the reference data derived from contour lines
(see 5.1).
The InSAR and stereo DEMs are very typical in both test sites.
While InSAR performs well in test site 1, where it reaches a
very good RMS error of 4.8m, while the RMS error drops in the
presence of steeper slopes of test site 2 to 14.9m (Tab. 2). This
behaviour proves the strong InSAR dependence on the
coherence in general and especially on terrain slopes, as
coherence in test site 2 is lower than in test site 1. In contrast to
SPOT DEMs, where outliers of more than 90m occur in both
sites, such extreme errors do not occur in the InSAR DEMs,
where maximum errors of 24m and 63m arise. The reasons for
this performance are the smooth terrain solution of LS
unwrapping and the height interpolation during the ground
range conversion process. Although outliers are avoided in
InSAR DEMs, the amount of errors greater than 20m is
significantly higher in test site 2, where InSAR performs worse,
than in the SPOT case of test site 1. This is due to the fact, that
LS unwrapping is capable to bridge only limited areas, but fails
to recover large decorrelated areas.
The SPOT stereo DEM performs robustly in both test sites. The
RMS in both types of terrain differs only by 1.4m. Also, the
correlation coefficient of test site 1 is with 0.7 a little higher
than that of site 2 (0.66), proving that accuracy follows the
correlation value also in the optical case. The robust
performance is also underlined by the amount of height
deviations greater than 20m. In test site 1, only 1.9% of all
values showed such an error, in test site 2, 4.3%. In both cases
almost all height errors (more than 99%) are below 40m. As a
conclusion, stereo DEM errors are in part extremely high, but
appear very localised and affect only their direct neighbours
(Fig. 4). As expected, the correlation coefficient shows the same
behaviour (Fig. 3).
Due to the relatively short repeat pass interval of both sensors
and the applied pre-processing, the mean correlation coefficient
is relatively high with values between 0.54 and 0.7 for both
sites and sensors. The cross-correlation shows distinct patterns
for both sensors. While it follows closely the shape of the
terrain in the SAR case, decreasing at the flanks of the hills and
thus varying regionally, it shows extreme local variations in the
optical case, where the correlation is corrupted at various spots.
As the low correlation values appear in different locations in
both InSAR and stereo-optical DEMs, the DEMs can be used
complementarily by the proposed fusion process, which is
demonstrated in the results of the fused DEMs.
Test site 1
Test site 2
Signed average: 1.0m
Average: 3.7m
RMS: 4.7m
Correlation: 0.61
Signed average: -2.9m
Average: 11.3m
RMS: 14.9
Correlation: 0.54
Signed average: 1.9m
Average: 6.0m
RMS: 7.9m
Correlation: 0.7
Signed average: -2.3m
Average: 6.8m
RMS: 9.3
Correlation: 0.66
Signed average: 1,4m
Average: 3.2m
RMS: 4.0m
Signed average: -2.2m
Average: 4.9m
RMS: 6.5m
Table 2. Single sensor and fused DEM errors (reference
DEM - generated DEM).
The RMS decreased after the data fusion in both test sites. In
test site 1, the RMS dropped to 4.0m after the fusion, a decrease
of 16% compared to the ERS and of 49% compared to the
SPOT DEM. In test site 2, the RMS dropped to 6.5m, a
decrease of 30% compared to the SPOT case and 56 % to the
ERS case (Tab. 2).
The fact that no error greater than 20m occurs in test site 1
proves the error sensitivity of the fusion procedure (Tab. 3). The
partially extreme outliers of the SPOT DEM are completely
rejected (Fig. 5) and only less erroneous values are fused with
the very accurate InSAR DEM of test site 1.