ırately
)vered
M and
le tree
SS Was
r map
reated,
z an F
asured
>d and
within
F ratio
If the
cemed
cliably
d; this
riation
nt was
nts, so
r map.
r is not
d cells
ted for
> cover
sample
Ising à
d cells.
nel for
y least
tions:
neters
DSM
cach
| tree
00);
e h is more than twice its estimated standard deviation,
since low values of h are usually caused by a poorly
matching tree map; and
e. h isless than 25 m to exclude unreasonably large
estimates that typically occur where tree patterns
coincide with terrain features.
The estimated parameters dg ...d4 are not used in the
subsequent process but are included in the model so that the
estimated tree offset is not overly influenced by correlations
between the orientation of the patch edge and the natural
gradient of the land.
2.4 Interpolation of tree offset across patches
The height offset estimates obtained from the least squares
estimation are only available for patch edges whereas heights
are required throughout the patches, so the height estimates
must be interpolated through the patches. This was achieved
using an early version of the multiscale adaptive smoothing
method (Gallant, 2011) that uses the variance information to
weight the estimates. This produces a continuous surface of tree
height offset both within and outside patches i.e. for both tree-
covered and non-tree-covered areas.
2.5 Subtraction of offset from DSM to produce DEM
The interpolated tree offset is multiplied by the smoothed tree
cover map to produce an estimate of the tree offset that can be
subtracted from the DSM to produce the bare-earth DEM:
2.6 Computational details
Most of the processing steps were implemented as AML macros
within ESRI Arc Workstation. The least squares fitting was
implemented as a C++ program.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
Due to the size of the SRTM DSM for Australia (about 40 GB
of single-precision floating point data) the processing was
performed on 1x1 degree tiles using sufficient overlaps in each
step that the results were consistent across tile boundaries. This
tiled processing approach also allowed for distributed
processing. At various stages of the project we used an 80-node
processing cluster, a multi-core server and a Condor distributed
processing system using idle desktop PCs.
The tree offset estimation was performed on the DSM product
after removal of stripes but without voids filled, so that the
interpolated data in voids would not influence height estimates.
After the subtraction of estimated tree offsets, voids were filled
and water bodies re-flattened as described in Read et al., (in
prep).
3. RESULTS AND DISCUSSION
Figure 1 shows an example of the removal of tree offsets in an
area where the method worked very effectively. The DSM
(centre panel of Figure 1) shows marked elevation offsets that
would seriously interfere with land surface attributes such as
slope and with flow paths computed from the DSM. After
treatment (DEM, right panel of Figure 1) there are very few
artefacts due to vegetation patterns. Both the regular patterns of
plantation forestry and the more erratic patterns of natural
vegetation in the valley (running from the eastern to southern
border of the area) are well represented in the tree cover map
and the height offset is accurately modelled by the algorithms.
Figure 2 shows the estimated tree height offset for the area
shown in Figure 1. The variation in estimated offset around the
patch edges where the offsets are measured contrasts with the
smooth variation within the patches where offsets are
interpolated from the edges. The lower estimated height within
the patches is due to variations in heights around the edges and
is probably an under-estimate: it is more likely that trees are at
Elevation (m
au 190
)
Figure 1. Landsat image (left), digital surface model (DSM, centre) and bare earth digital elevation model (DEM, right) near Milltown,
Victoria, 141.78°E 38.08°S
