Figure 2. Estimated tree height offset for the arca shown in 
Figure 1. 
least as high within the core of the patch compared to the edges. 
The impact of these errors is relatively minor and the method 
adequately deals with the most troublesome aspect of the 
artefacts, the abrupt changes in height at patch edges. 
The techniques described here were applied to 813 1°x1° tiles 
covering Australia. Effective removal of vegetation offsets, 
assessed by visual examination of DEM, was achieved for about 
90% of the vegetated area of the Australian continent. The 
remaining areas contain offsets that were untreated or only 
partially removed, or areas where the offsets were over- 
estimated. Most of these defects are related to the tree cover 
mapping and have several causes including: 
e the mapping could not distinguish between trees and 
other ground cover; in some cases areas of trees are 
missed while in other cases areas not covered by trees 
are classified as tree-covered 
e there were significant changes of tree cover in the 
time between the imagery capture for the mapping 
and the SRTM radar mission 
  
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 
Figure 3. Landsat image (left), DSM (centre) and DEM (right) near Nelson, Victoria, 141.08?E 37.98°S 
Some of the defects are due to the algorithms, particularly over- 
estimation of tree height offset where vegetation and terrain 
patterns coincide (such as tree covered rises in cleared 
agricultural land) or in areas of low vegetation height where the 
constraints on acceptable height estimates may have produced 
biased estimates by omitting small but correct height offsets. 
Coincidence of vegetation and terrain patterns can also result in 
under-estimated offsets, for example in riparian forests filling 
inset floodplains. The interpolation of height offsets across 
large areas of continuously wooded landscape also appears to 
have resulted in under-estimation of tree height offset in much 
of the interior of the patches. 
Figure 3 shows an area where some areas affected by tree 
offsets have not been effectively removed because of differences 
between the tree cover map and the patterns of offsets in the 
SRTM DSM, presumably due to changes in tree cover between 
the Landsat and SRTM acquisition dates. The regular patchy 
nature of the tree cover is due to plantation forest harvesting 
and replanting. Figure 3 also illustrates over-estimation of tree 
height offsets around the river. The river is incised, bordered by 
cliffs nearly 20 m high, and the algorithm has been unable to 
separate the co-incident effects of sudden terrain change and 
tree cover change thus interpreting the cliffs as tree heights and 
substantially over-estimating the tree height offset. The removal 
of this apparent offset has largely removed the cliffs and 
produced land heights beside the river comparable to the river 
surface height. 
Improvements in the quality of the derived bare-earth DEM 
could readily be achieved by correcting parts of the tree cover 
map that did not correspond well with the offsets in the SRTM 
DSM. It is also likely that the algorithms could be enhanced to 
provide more accurate results. 
3.1 Applications within Australia 
The bare-carth DEM produced from the SRTM DSM is still 
quite noisy and lacks the hydrological connectivity needed by 
some applications. The DEM for Australia has been further 
processed to produce a smoothed (DEM-S) product and a 
hydrologically enforced product (DEM-H). From these two 
products a suite of commonly used terrain attributes have been 
derived, which are now being used for ecological, hydrological 
and geomorphological purposes. The stream networks and 
catchments derived from DEM-H will underpin the ongoing 
refinement of the Australian Hydrological Geospatial Fabric 
(m)