woodlands are also problematic: many of Australia’s most
significant rivers run through very low relief terrain and are
typically bordered by tall trees while the surrounding landscape
is either cleared for agriculture or is naturally devoid of dense
tree cover. The rivers therefore appear in the SRTM DSM as
raised ridges rather than channels. Collectively these various
tree offsets prevent effective use of the DSM for most analytical
purposes.
Manual, or manually assisted, workflows for removing offsets
from DSMs exist within commercial organisations but were
considered too expensive to use across an entire continent.
Fully automated methods were therefore developed to remove
the offsets due to trees. The methods produced effective results
over most of the vegetated parts of the continent (about half of
the area of Australia required some removal of tree offsets).
Some manual effort was required to select the best tree cover
map for use in different areas and to evaluate the results but the
processing itself is fully automated. Offsets due to buildings and
other structures were not treated.
The opportunity now exists for these methods to be applied
elsewhere, potentially producing a global bare-carth DEM
suitable for most analytical applications. The estimated tree
height offset is also likely to be valuable as a high-resolution
surrogate for biomass. With suitable adaptation, the methods
could also be adapted to other DSMs such as TANDEM-X to
fully automate the production of bare-earth DEMs.
2. METHODS
The existence of tree offsets in the SRTM data was an expected
characteristic of the data, and several studies have investigated
the relationship between tree height and the induced offset in
SRTM elevations where a reference DEM was available (e.g.
Kellndorfer et a/., 2004; Walker et al., 2007). At the time this
work was commenced there were no published examples of
estimation and removal of the offset where a reference bare-
earth DEM was not available, and even now the authors are
aware of only one other similar effort (Kóthe and Bock, 2009).
The method described here relies on an independent map of tree
cover so that the areas requiring treatment can be defined. This
map must be compatible with the DSM resolution and represent
the conditions at the time of the DSM acquisition (February
2000 in the case of SRTM). Land cover maps derived from
Landsat TM at about 30 m resolution are the obvious choice,
and in Australia we were fortunate to have several such maps
derived from Landsat. These maps were produced using
different processes and from imagery at different dates so a
manual selection process was applied to choose the best
representation of the visible SRTM offsets in different parts of
the country.
The main steps in the tree offset removal process are:
1. Adjustment of the tree cover patches to match edges
in the SRTM DSM;
Characterisation of the SRTM DSM to patch edges;
Estimation of tree offset at patch edges;
Interpolation of offset to the interior of patches; and
Subtraction of the tree offset surface from the DSM.
ud wt
A full description of the processing steps and the compilation of
the tree cover map will be provided in Gallant ef al (in prep).
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
2.1 Adjustment of the tree cover map
The correct estimation of tree offset relies on accurately
identifying the location of the transition from non-tree-covered
to tree-covered terrain. Mis-registration between the DSM and
the tree cover map will result in under-estimates of the tree
offset and ineffective removal of the offsets, so a process was
developed to adjust the edges of patches in the tree cover map
to optimise the match between the map and the DSM.
Several modified versions of the tree cover map were created,
and the effectiveness of each version was assessed using an F
ratio in a neighbourhood around each grid cell that measured
the difference in average height between the tree-covered and
non-tree-covered area, relative to the variation in height within
both areas. For each cell, the version with the highest F ratio
was chosen as the best representation of the patch edge. If the
best F ratio was smaller than a minimum value, this was deemed
to indicate that there was insufficient information to reliably
adjust the patch edges and the original map was used; this
mostly occurred in higher relief areas where the terrain variation
masks the offset due to vegetation. The range of adjustment was
limited to two grid cells to prevent unreliable adjustments, so
the method cannot correct for large errors in the tree cover map.
2.2 Characterisation of edge response
The response of the SRTM DSM to changes in tree cover is not
sharp but transitions smoothly over a distance of 3-4 grid cells
(around 100 m). This smooth transition must be accounted for
in the correction of offsets to avoid artefacts around tree cover
patches.
The response to edges was assessed over a large (2105) sample
of clearly defined edges in flat terrain and modelled using a
Gaussian smoothing kernel with a length scale of 1.4 grid cells.
The adjusted tree cover map was smoothed using this kernel for
use in the remainder of the processing.
2.3 Estimation of tree offset at patch edges
The tree offset was estimated near every patch edge by least
squares estimation based on a model of local height variations:
ZDSM (de, $) = a, + a, + a, y + a, + hmç +E (1)
where Znsy - DSM elevation
&, dy = location relative to target cell
ag ...d4 = local terrain height and slope parameters
h = tree height offset
mg = tree cover map after smoothing with
Gaussian kernel
€ = random noise
The parameters dy ...a; and A are fitted to values of Z psy
and mg from a circular window of 5 cell radius around each
target cell using linear least squares fitting. The estimated tree
height / is retained if:
e the model fits well enough (defined by Y « 200);
e. the estimated variance of À is less than 3 m;
