EM.
aitment
to 30m
ndicate
area in
oduced
ry low
esulting
ements.
1e users
level 2
mances:
ality of
Remote
ation of
search
ed from
ence on
ibsidies.
acy of a
st 2010,
c frame
ongress,
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
REMOVAL OF TREE OFFSETS FROM SRTM AND OTHER
DIGITAL SURFACE MODELS
J. C. Gallant^, A. M. Read, T. I. Dowling*
* CSIRO Land and Water, Black Mountain Laboratories, Clunies Ross St, Acton ACT 2602 Australia
(John.Gallant(g)csiro.au (corresponding author), Arthur.Read(g)csiro.au, Trevor. Dowling@csiro.au)
KEY WORDS: DEM/DTM; Radar; Correction; Vegetation; Hydrology; Geomorphology
ABSTRACT:
The recently completed 1 second Digital Elevation Models (DEMs) for Australia are based on the 1 second Shuttle Radar
Topographic Mission (SRTM) elevation data. The SRTM data was corrected by removing voids, striping, tree offsets and random
noise and finally by integrating mapped drainage lines. This paper describes the removal of the tree offsets, which was a crucial step
in the production of a credible bare-earth elevation model and was one of the most technically challenging aspects of the project, and
the possible application of the methods to other digital surface model (DSM) sources.
Methods for the removal of tree offsets rely on maps of tree presence/absence from sources such as remotely-sensed imagery, and the
height offsets are computed from the DEM at the boundaries of tree patches. Tree offsets over most of Australia were successfully
removed, but were underestimated in areas of extensive forest cover and poorly estimated where the mapping of tree patches did not
match the patterns of offsets in the SRTM elevations.
The tree offset removal methods could be applied to the near-global SRTM DSM to produce a near-global bare-earth product,
provided that a suitable map of tree presence or density can be compiled from satellite remote sensing and other sources. The process
could be improved by using supplementary tree-height information from ICESat or other sources.
High resolution global DEMs other than SRTM are becoming available, notably ASTER GDEM and TANDEM-X. Both those
products are subject to offsets due to vegetation in the same way as SRTM. The tree offset removal methods developed for SRTM
could be adapted to the characteristics of these and other DSMs to provide a largely automated processing system to derive bare-
earth DEMs from new sources.
1. INTRODUCTION
The release of the Shuttle Radar Topographic Mission (SRTM;
Farr et al., 2007) data for Australia in 2005 heralded a
significant step forward in resolution, detail and consistency for
elevation models over that continent. Previous continental
DEMs had been prepared from topographic mapping at scales
of 1:2.5M to 1:100k with the most recent being the Geodata
9second DEM version 3.0 (http://ga.gov.au/topographic-
mapping/digital-clevation-data.html). The impact of this
improvement in resolution is enormous. Most hillslopes are on
the order of 200 m long so changing from 250 m to 30m
resolution brings the ability to resolve the basic ridge-valley
structure of the land surface that is crucial to understanding the
flow of water and materials through the landscape and
everything that follows from that, in particular the patterns of
soil depth and properties and the distribution of vegetation
communities.
While the improved resolution provided obvious benefits, the
defects in the SRTM data were also readily apparent: in contrast
to the smooth, complete and hydrologically enforced 9 second
DEM the SRTM data was noisy, had missing data (voids),
contained stripes, was subject to significant offsets in vegetated
areas and did not capture river channels. In short, it is a noisy
digital surface model (DSM) rather than a bare-carth model
(DEM or DTM). This led to a collaborative project between
CSIRO, the Bureau of Meteorology, Australian National
University and Geoscience Australia to deal with each of those
defects extending from 2005 through 2011. Access to the 1
second SRTM data (which is not routinely released for
territories outside the USA) was provided by DIGO, Australia's
defence mapping agency.
Widespread diagonal striping with a wavelength of about 800 m
was detected and treated using a 2-dimensional Fourier
transform in a custom-built tool that allowed manual
identification of the striping frequency and orientation in
Ya xX Ya degree tiles (Read et al., in prep). Voids were removed
using a modification of the Delta Surface Fill method (Grohman
et al., 2006), with the 9 second Geodata DEM providing infill
data (Read ef al., in prep). Tree offsets were removed using the
methods described in this paper. Random noise was reduced
using an adaptive smoothing method (Gallant, 2011) that
smooths to a greater or lesser degree in response to the noise
amplitude (estimated from the SRTM DSM) and the local relief.
Hydrological connectivity was enforced with a modified version
of the ANUDEM program (Hutchinson, 2011) using mapped
1:250k stream lines modified to fit the SRTM DSM in higher
relief areas (Dowling ef al., in prep). The resulting set of four
products (a cleaned DSM, bare-earth DEM, smoothed DEM-S
and hydrologically enforced DEM-H) are now publicly
available through Geoscience Australia's National Elevation
Data Framework Portal (http://nedf.ga.gov.au) at both 1 second
and 3 second resolutions (except for the DSM, which is only
available for government use at the 1 second resolution, and
DEM-H, which is only available at 1 second resolution due to
the difficulty of generalising while retaining the necessary
hydrological connectivity).
This paper describes the methods developed to remove tree
offsets from the SRTM DSM, which was a crucial step in the
production of a credible bare-earth DEM for Australia. These
tree offsets are clearly visible in the SRTM DSM in agricultural
lands and managed forests where there are abrupt transitions
between cleared and tree-covered areas. Riparian forests and
