In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
Figure 3: Overview of the test site “Burgau”. This topographic
map shows regions of forest in green color.
Name
Look angle
GSD
Date
MGD_ascl
22.3°
1.25m
2009-07-28
MGD_asc2
37.2°
0.75m
2009-08-02
MGD_asc3
48.5°
0.75m
2009-08-07
MGD.dscl
21.3°
1.25m
2009-07-30
MGD_dsc2
36.5°
0.75m
2009-08-05
MGD_dsc3
48.0°
0.75m
2009-07-31
Table 1 : Detailed parameters for the Spotlight images.
Forest Segmentation. The InSAR coherence used in the
segmentation process is derived from a TerraSAR-X single look
complex (SSC) InSAR pair (see Table 2). These images were
ordered as dual-polarization products (HH,VV) with science
orbit accuracy and were acquired in March 2008.
Name
Look angle
Date
SSC-dscl
36.0°
2008-03-17
SSC_dsc2
36.0°
2008-03-28
Table 2: Detailed parameters for the Spotlight InSAR pair.
Reference Data. To enable quantitative evaluations LiDAR data
is used as ground truth information. The LiDAR reference
data covers four measurements per square meter, which are
processed to highly accurate DSMs and DTMs. While the DSMs
are automatically extracted, the DTMs are semi-automatically
generated by classifying regions of vegetation, building, bridges
and other man-made structures.
To evaluate the DSMs derived using TerraSAR-X multi-image
radargrammetry, 30 regions on bare ground and 70 regions in
forest are selected. The average residual height error over such
regions describes the DSM quality. Over forest, the average
canopy height underestimation r is extracted and used to correct
the canopy height.
To evaluate the image segmentation quality a ground truth
reference mask is derived using LiDAR data. The lm GSD
LiDAR CHM is filtered with an order-statistic filter of size 7x7
and order 37, i.e. the 75^ percentile. The CHM is then down
sampled to a GSD of 5m using a 5 x 5 average resampling.
Next, pixels with a height larger than 8 meters are considered as
forest regions and small regions are filled to eliminate noise.
4 RESULTS AND DISCUSSION
Multi-Image DSM Generation. For visual interpretation some
detailed results are shown in Figure 4. All results are given with
a GSD of 2 meters in UTM33 projection. Figure 4(a) shows
the TerraSAR-X DSM, (b) the LiDAR reference DTM, (c) the
TerraSAR-X and LiDAR based CHM, (d) the pure LiDAR CHM,
(e) the TerraSAR-X based height error (i.e. the ground truth
LiDAR DSM subtracted from the TerraSAR-X derived DSM)
and (d) a topographic map. The TerraSAR-X CHM corresponds
visually very well to the LiDAR CHM and to the topographic
map, however the TerraSAR-X DSM is too low over forest, as
seen in Figure 4(e). Regions of bluish color indicate height un
derestimations and such regions are placed in forests or on for
est borders. The quantitative accuracy analysis is listed in Ta
ble 3. Heights over bare ground are reconstructed with very high
accuracy (mean value below 20 cm and standard deviation of
about 2 meters). The canopy height however is systematically
underestimated by approximately 27.5% for this test site. In our
previous work we estimated an average underestimation using
Spotlight and Stripmap imagery and multiple scenes which was
26.6%±1.4% (Perko et al., 2010). When correcting the height
bias with this learned value the reconstruction over forest be
comes a lot better. In particular it decreases to a residual height
error of 20 cm, like on bare ground (see Table 3 bottom). Figure 5
and Figure 6 show detailed analyses of the canopy height under
estimation w.r.t. the canopy height before and after the discussed
correction, clarifying that the underestimation over forest can be
corrected for this test site.
bare ground
forest
ß [m]
a [m]
/i [m]
cr [m]
r [%]
asc
0.07
2.04
-5.81
1.87
28.5
dsc
0.18
1.90
-5.45
1.83
26.4
forest after height correction
asc
0.22
1.87
-1.16
dsc
0.08
1.83
-0.42
Table 3: Detailed 3D height analysis of the TerraSAR-X derived
DSMs.
Forest Segmentation. The features used for forest segmenta
tion are shown in Figure 7. Obviously, the most important in
formation for the segmentation are the InSAR coherence and the
canopy height model. The confusion matrix in Table 4 reveals
that 90% of pixels (here one pixel has a GSD of 5 meters) are
correctly classified. About 8% of non-forest regions are classi
fied incorrectly as forest. This especially happens in small forest
clearances which are not seen due to the slant range SAR geome
try or which result from image matching failures. The 2% of pix
els classified wrongly as non-forest are mainly small forest stands
where image matching is unsuccessful and thus such regions get
interpolated. In addition it should be noted that this evaluation is
relative to the LiDAR ground truth segmentation. Therefore, the
achieved accuracy is most likely above the 90% since some arti
facts exist in the LiDAR model. For instance some power supply
lines are classified as forest. In comparison to the state-of-the-art
classification of TerraSAR-X data in (Breidenbach et al., 2009)
the proposed method performs very well. On first glance their
method also reaches a classification accuracy of 90%. However,
the evaluation is based on image blocks with 20 x 20 m 2 . When
reducing the GSD to 5 meters, like in our study, the classification
accuracy of (Breidenbach et al., 2009) drops to 72.5%. Obvi
ously, our method performs better as a diversity of information
like canopy height model, coherence or texture descriptors is in
corporated into the classification process.