Technical ReportPDF Available

A new method for the removal of off-terrain objects from LiDAR-derived raster surface models


Abstract and Figures

A new algorithm, named ROTO, is presented for creating bare-earth digital elevation models (DEMs) from input raster digital surface models (DSMs). The algorithm can remove off-terrain objects (OTOs) from DSMs derived from light-detection and ranging (LiDAR) and other fine-resolution topographic data sources. The accuracy of the method was tested using three LiDAR DSMs of sites possessing a variety of urban, agricultural, and forest land-covers and a range of topographies. The ROTO-derived bare-earth surfaces were compared against reference data sets created by applying ground-separation techniques to the raw LiDAR point clouds. The analysis demonstrated that ROTO calculates accurate bare-earth surfaces and mapped OTOs in agricultural, forested, and urban settings, with measured accuracies between 0.159 m and 0.499 m at the 95% confidence level. OTOs mapped using the new algorithm were found to correspond closely to the mapped features from the reference data set, with kappa index values ranging from 0.896 to 0.962. This study showed that similar accuracy can be attained during OTO removal and mapping when end-users only have access to interpolated raster LiDAR data products rather than the point-cloud source data.
Content may be subject to copyright.
A new method for the removal of off-terrain objects from LiDAR-derived raster sur-
face models
John B. Lindsay*
Department of Geography, The University of Guelph, Guelph, Canada
* Corresponding author,, E-mail:; Phone: 519-824-4120 ext.
56074, Department of Geography, University of Guelph, 50 Stone Road East, Guelph, N1G 2W1, Canada
Note:This is a working paper intended to serve as a reference for the Remove Off-terrain Objects (ROTO) tool found in
Whitebox GAT. There are no plans to submit this work for future publication and so this document serves as the main
reference for the algorithm.
A new algorithm, named ROTO, is presented for creating bare-earth digital elevation models (DEMs) from input
raster digital surface models (DSMs). The algorithm can remove off-terrain objects (OTOs) from DSMs derived from
light-detection and ranging (LiDAR) and other fine-resolution topographic data sources. The accuracy of the method
was tested using three LiDAR DSMs of sites possessing a variety of urban, agricultural, and forest land-covers and a
range of topographies. The ROTO-derived bare-earth surfaces were compared against reference data sets created
by applying ground-separation techniques to the raw LiDAR point clouds. The analysis demonstrated that ROTO cal-
culates accurate bare-earth surfaces and mapped OTOs in agricultural, forested, and urban settings, with measured
accuracies between 0.159 m and 0.499 m at the 95% confidence level. OTOs mapped using the new algorithm were
found to correspond closely to the mapped features from the reference data set, with kappa index values ranging
from 0.896 to 0.962. This study showed that similar accuracy can be attained during OTO removal and mapping
when end-users only have access to interpolated raster LiDAR data products rather than the point-cloud source
Keywords: LiDAR, DEM, off-terrain objects, data processing
1. Introduction
Airborne light detection and ranging (LiDAR) data, sometimes called airborne laser scanner (ALS) data, are increas-
ingly applied to activities involving detailed topographic mapping because of their fine spatial resolution and high
accuracy and the relatively low cost per point measurement (Nelson et al. 2009). LiDAR based topographic mapping
also has the advantage over competing technologies that it is able to partially penetrate forest canopies and other
vegetation cover to measure the underlying ground surface (Hodgson et al. 2003). These characteristics have led
some authors to assert that LiDAR has become the new standard for the creation of DEMs (Liu, 2008).
Airborne LiDAR systems acquire the horizontal and vertical position of reflective surfaces beneath an aircraft-
mounted laser scanner. Each measurement represents an (x, y, z) point within the raw, unstructured LiDAR point
cloud, which typically contains many millions of individual measurements. Each point in a data cloud therefore
records the location of a return from a laser pulse that was reflected off of either the ground surface or an object
situated between the LiDAR system and the ground. Non-ground points may be reflected off of trees, buildings,
fences, power lines or any other object above the ground surface, known as an off-terrain object (OTO). Many
applications of digital topographic data require a DSM with OTOs that have been removed, commonly referred to
as a bare-earth DEM (Liu, 2008). (Note, the term DSM is used in this paper to refer to a surface model that includes
OTO elevations, while DEM is used to denote a ground-surface model.) For example, OTOs must be removed from
LiDAR DEMs that are used for modeling surface drainage or sediment transport. OTOs are also separated from the
underlying terrain surface to map buildings and other anthropogenic structures (Rottensteiner and Briese 2002,
Sohn and Dowman 2001, Tarsha-Kurdi et al. 2006) and to study forest canopy characteristics (Chen et al. 2006,
Drake et al. 2002, Lim et al. 2003, Popescu et al. 2003, Suárez et al. 2005, Zimble et al. 2003).
A single laser pulse can result in multiple return points because the laser’s footprint can intersect more than one
surface as it approaches the ground. For example, multiple returns can be measured if a pulse hits a surface that
is smaller than the area of the footprint (e.g. a tree branch) or an object’s edge. Many LiDAR systems are capable of
recording the first and last returns and some systems are capable of recording five or more return points for each
laser pulse (Flood 2001). The last-return point of a pulse is often assumed to correspond to the ground elevation,
although this is not necessarily always the case. In areas of forest cover, a type of semi-transparent OTO, the last-
return of a laser pulse is nearer the ground surface but these points can also correspond to tree trunks or woody
debris lying on the ground. For opaque OTOs such as buildings, there is usually only one return, except where the
laser footprint intersects the building edge. As such, a raster DSM that is created by interpolating last-return LiDAR
points is unsuitable for use as a bare-earth DEM and additional processing is often needed to remove OTOs.
In this paper a new algorithm is presented for removing OTOs from LiDAR DSMs for the creation of bare-earth
DEMs. This algorithm is therefore also useful for OTO mapping applications, such as building extraction and canopy
modeling. The ability of this new algorithm to create bare-earth DEMs and to identify OTOs is evaluated under a
variety of land covers at three study sites. Although the approach described in this study is suited to application with
LiDAR data, it can also be used for creating bare-earth DEMs from any fine-resolution elevation data that contains
these features.
2. Related Work
Several algorithms for removing OTOs from LiDAR data have been developed over the last two decades (Arefi and
Hahn 2005, Kobler et al. 2007; Sithole and Vosselman 2001, Sohn and Dowman 2008, Suárez et al. 2005, Tarsha-Kurdi
et al. 2006, Zhang et al. 2003; Evans and Hudak 2007). This has, however, been identified as a critical area for further
development and improvement (Liu 2008, Meng et al. 2010). There are two general approaches to the creation of a
bare-earth DEM from LiDAR data. The first approach involves classifying ground points within the raw LiDAR point
cloud and then excluding non-ground points from the process of interpolating the raster DEM. The second method
begins with a raster DSM that has been interpolated from the raw LiDAR data. The interpolation usually includes
the last-return points only. OTOs are then identified in the DEM, often based on a morphological characteristic that
is in contrast to the surrounding landscape. OTOs are then removed from the elevation model and new surfaces
are interpolated in their place.
Lui (2008) favours algorithms that are based on the first approach ( i.e. ground-point separation from LiDAR point
clouds) arguing that the initial interpolation process of the second approach results in a loss of information and
introduces error. If the data end user has access to the LiDAR point cloud, ground point separation techniques are
undoubtedly preferable. However, end users frequently only have access to the interpolated last-return raster data
product. The raw point cloud is sometimes treated as an intermediate data set and is either not supplied to end
users by data providers (end-users sometimes have to request the raw point cloud and may incur an additional
cost) or else the point cloud is discarded because of the additional storage requirements. Analyzing raster DSMs
instead of the raw point cloud also has the benefit of allowing for the use of the extensive and efficient techniques of
digital image processing. Algorithm performance is often a major challenge when working with massive LiDAR point
clouds and can be a limiting factor to their application (Guan and Wu 2010). Thus, while ground-point separation
techniques that are applied to raw LiDAR point cloud have advantages, there are also compelling reasons for the
development of DSM-based methods for separating OTOs from the underlying ground surface.
3. ROTO Algorithm Description
Figure 1 describes the workflow of the new algorithm for creating a bare-earth DEM, referred to as remove off-
terrain objects (ROTO). The method begins by performing a white top-hat transform (WTT) to remove the large-scale
topographic variation within the input DSM. A WTT involves differencing the mathematical morphological operation
known as an opening from the original DSM. In mathematical morphology, an opening is a dilation operation per-
formed on a previously eroded image, which is equivalent to treating the input DSM with a minimum filter followed
by a maximum filter, using the same filter kernel size. This approach has been used previously for ground point sep-
aration algorithm applied to LiDAR point clouds (Li et al. 2014). The erosion and dilation operations require the user
to input a filter kernel size parameter. Because the WTT involves filtering, this step of the algorithm’s workflow has
the potential to be the most computationally demanding component, particularly because relative large kernel sizes
are often required. The computational efficiency of these operators can be significantly improved by storing the
redundant information within the overlapping portions of the kernels for neighbouring grid cells and swapping out
only the data within the leading and trailing columns as the kernel moves along a row (Huang et al. 1979; Sizintsev
et al. 2008). This filtering approach was adopted for the WTT subroutine within the ROTO implementation.
Figure 1: Workflow of the ROTO algorithm.
Grid cells for which their WTT value is equal to zero (i.e. their elevation in the input DSM lies on or near the estimated
ground elevation) are flagged and placed into a queue. A queue is a common first-in-first-out (FIFO) data storage
structure. The queue serves to provide seed points for a region-growing operation, in which the algorithm iterates,
popping the front grid cell from the queue and searching its neighbouring grid cells for possible new entries into
the queue. Neighbouring grid cells are entered into the queue if the slope between the neighbour and the centre
cell, calculated based on the residual elevations of the WTT, is less than a user-specified slope threshold parameter
(Figure 1). This iterative process continues until the queue is empty. Grid cells that pass through the queue have their
elevations (i.e. not WTT-derived residual elevations) directly entered into the output bare-earth DEM. Any cells that
do not pass through the queue belong to areas within the DSM that are not connected to the underlying ground
surface due to the presence of an enclosing steep edge. In addition to being enclosed by steep edges, the use
of the WTT-derived residual elevations in the region growing operation ensures that cells that are not discovered
during the operation are also in an elevated position above the underlying terrain, an inherent characteristic of OTOs.
Finally, new elevation values are estimated for each OTO cell using an inverse-distance weighting (IDW) interpolation
scheme, based on the input elevations of nearby non-OTO grid cells. The search distance used to locate nearby
points is equivalent to the filter distance used in the WTT operation.
The ROTO algorithm has been implemented as a tool within the whitebox-tools geospatial analysis library and
was developed using the Rust programming language. The tool can read and write a variety of common
raster data formats including. Native binary executable files, compiled for MS Windows, MacOS, and Linux
operating systems, as well as the program’s source-code, are all distributed under an open-source license
( The ROTO executable file can be called through scripting
automation, enabling the tool to be readily embedded in existing geographical information system (GIS) software.
As an example, the tool has been integrated in the open-source software Whitebox Geospatial Analysis Tools
(Lindsay 2016) as the plugin tool Remove Off-Terrain Objects, which includes a user-interface and full integration with
the GIS.
4. Method
4.1 Test sites and data set descriptions
Three LiDAR data sets were used to evaluate the performance of the ROTO algorithm. The three data sets were
interpolated from LiDAR point clouds (LAS files) to 1 m grid resolution DEMs using a nearest-neighbour interpolation
routine. Last-return and only-return points were used in the interpolation process. The average density of LiDAR
point measurements in the three study sites were 2.06 points/m2, 2.37 points/m2, and 0.84 points/m2respectively.
The relatively lower density of last-return point measurements at the third study site reflected the greater coverage of
dense forest within the area. The three test sites were chosen because they exhibited a range of urban, agricultural,
and forested land-covers in addition to varying topographic conditions.
Site 1 (Figure 2) is 6 km2and includes a portion of the town of Blenheim located in south-western Ontario, Canada.
The area possesses a mixture of residential, commercial, and industrial land-uses, which is apparent in the LiDAR
intensity image (Figure 2A). Buildings are clearly the dominant OTO present in the 2,000 × 3,000 (rows by columns)
last-return interpolated DEM (Figure 2B). Several larger buildings are present in the site, including two schools, an
arena, and a retirement complex. The residuals of numerous trees are also apparent in the DEM, particularly associ-
ated with the residential areas and several woodlots in the northern half of the site (Figure 2B). The site is relatively
flat, with an overall vertical relief of approximately 20 m. A low ridge dominates the local topography and occupies
the central area.
Site 2 is a 4 km2area located within the town of Picton, in eastern Ontario, Canada. Large parts of the site are under
heavy deciduous forest cover, which is apparent in both the LiDAR intensity image (Figure 3A) and the shaded-relief
image derived from the 2,000 × 2,000 last-return DEM (Figure 3B). The LiDAR data were acquired during leaf-off
conditions. The forest cover extends from a centrally located, steep-sided river valley to a plateau located along
the southern and eastern edges of the site. Numerous small to medium sized buildings are present throughout the
western and northern sections of the site, while the largest buildings are associated with the regional airport located
atop the plateau near the bottom central region. The site contained moderate overall relief, with 76 m of elevation
Figure 2: Site 1 LiDAR intensity image (A) and last-return LiDAR DEM (B).
difference between the plateau and the adjoining valley bottom. The western and northern sections of the site are
dominated by residential, commercial, and light-industrial land-uses.
Site 3 is a 9 km2area within Jay State Forest, near Montgomery, VT, USA (Figure 4). Thick and nearly continuous
mixed forest covers the region (Figure 4A), which is contained within the northern Green Mountains. The LiDAR data
were collected during autumn 2008 and spring 2009, during leaf-off conditions. The Jay State Forest site has the
greatest relief of the three study areas, with 634 m of elevation difference between the mountainous ridge tops and
the Black Falls Brook valley bottom, which drains toward the south-western corner of the site. Jay Peak is located
south-east of the study area. The interpolated 2,765 × 3,282 last-return DEM exhibits considerable residual forest
cover that obscures the headwater drainage features in many locations (Figure 4B). There are no obvious buildings
or other structures within the site.
4.2 Algorithm Evaluation
ROTO’s performance was evaluated by comparing the resulting bare-earth DEMs with reference DEMs interpolated
from ground-separated LiDAR point clouds. The ground separation procedure involved applying an initial adap-
tive slope-based filter, similar to the method proposed by Sithole and Vosselman (2001), followed by a point-cloud
segmentation with manual re-classification of ground-surface segments. The slope-based ground separation was
effective at removing vegetation and smaller buildings and the segmentation procedure was used to remove larger
buildings in the Site 1 and Site 2 data sets. The ground-separated point clouds were then interpolated to raster
DEMs using the same method used to create the last-return DEMs to which the OTO algorithm was applied. These
DEMs served as the reference data set for an accuracy assessment of DEM heights, which included estimations of
the mean vertical error, root-mean-square-error (RMSE), and the accuracy at a 95% confidence level. Additionally,
two-class images of ’ground-level’ and ’OTO’ were created for both the reference data and the new-algorithm treated
DEM. OTO grid cells were identified as cells for which the bare-earth DEM height was less than the corresponding
height in the original last-return DSM. These class maps were used to assess the positional accuracy of OTO maps
based on Kappa analyses, a technique commonly used to perform accuracy assessment for classified data (Liu and
Mason 2009).
5. Results
Figure 5A shows the bare-earth DEM that was created by applying the ROTO algorithm to the Blenheim site last-
return LiDAR DEM. The algorithm removed all of the buildings and trees from the original DEM, leaving the gently
undulating natural topography. The only evidence of buildings in the bare-earth DEM was in the surfaces associated
with the level land gradings (i.e. the constructed platforms on which buildings are situated) of the larger buildings
in the site. The ROTO-applied DEM and reference DEM were nearly visually indistinguishable. The accuracy of the
Site 1 ROTO-applied DEM was the highest of the three sites, with an RMSE of the of 0.081 m (Table 1). The OTO map
(Figure 5B) shows the heights of each OTO in the study site. The Site 1 Kappa coefficient was 0.962 and therefore
the level of agreement between the OTO maps produced by the classification procedure and ROTO algorithm was
considerably higher than would be expected by random chance.
Table 1: Accuracy assessment of ROTO-applied DEMs compared with bare-earth DEMs generated from ground-
separated LiDAR point clouds.
Site 1 Site 2 Site 3
Mean vertical error (m) -0.007 0.023 -0.015
Site 1 Site 2 Site 3
RMSE (m) 0.081 0.228 0.255
Accuracy at 95% confidence level (m) 0.159 0.447 0.499
Kappa coefficient 0.962 0.934 0.896
The RMSE (Table 1) for the Site 2 ROTO-applied bare-earth DEM (Figures 6A) was 0.228 m (0.023 m mean vertical
error; 0.447 m accuracy at the 95% confidence level). The ROTO algorithm exposed the underlying terrain of the river
valley and plateau, leaving little evidence of the dense forest covering obscuring the ground in these areas within the
original last-return DEM (Figure 3B). The buildings located within the more heavily developed north-western portions
of Site 2 were largely removed using the algorithm (Figures 6A and 6B). The level of agreement in the placement of
OTOs between the ROTO and reference data mapped features was 93.4% higher than would be expected due to
random chance (Table 1). Several higher gradings within the site, including an abandoned quarter-mile track atop
the plateau, were preserved in the bare-earth DEM due to their gently sloped edges.
The accuracy assessment showed that the ROTO algorithm performed worse on the steeply sloped and densely
forested Site 3 data set (Table 1) than the other two study sites. However, the measured RMSE of 0.255 m in Site 3
was of a comparable magnitude to the vertical accuracy of many airborne LiDAR datasets. With no buildings located
within the site, all of the mapped OTOs (Figure 7B) were associated with residual forest cover predominately located
along the ridge tops and towards the valley bottom. The Kappa coefficient of 0.896 demonstrated a high level of
agreement between the OTOs mapped using the ROTO algorithm and the point-cloud based ground-separation
Visual inspection showed that ROTO did well creating bare-earth surfaces in all of the agricultural, residential, com-
mercial/industrial, and densely forested areas of the three test sites. Figure 8 is a comparison of hillshade images
derived from the DSMs, ROTO-derived bare-earth DEMs, and the reference data for portions of Site 3 and Site 2. Al-
though the Black Falls Brook valley bottom (Figure 8A) is completely obscured by residual forest cover in the original
DSM, ROTO was able to remove nearly all of this ground cover to reveal the underlying braided river topography
(Figure 8B). Furthermore, there was little distinguishable difference between the ROTO processed DEM and the ref-
erence data (Figure 8C) in this area. The Site 2 detailed image (Figure 8D) shows a portion of Picton’s residential
and commercial area and the adjacent forested parkland to the east. ROTO performed well in removing most of
the buildings in the site (Figure 8E) and provided a very similar bare-earth surface to that of the point-cloud derived
reference data (Figure 8F). The bare-earth surfaces showed more significant differences within the densely forested
and steep valley side near the center and bottom of Figures 8E and 8F.
The slope threshold and filter kernel size parameters were set to values that resulted in the removal of all apparent
OTOs in the three study site LiDAR DSMs. The slope threshold was set to 12.0° for Site 1 while a 15.0° threshold was
used for Sites 2 and 3. The kernels were set to 75 cells, 49 cells, and 11 cells for Sites 1, 2, and 3 respectively. The
kernel size that is needed to remove all OTOs from a DSM is a function of the grid resolution and the most extensive
OTO in the scene. Kernel size and the abundance and complexity of OTOs were found to affect the processing time to
a greater extent than the raster size (i.e. number of grid cells). On a computer with a 3.3 GHz processor and 16 GB of
memory the processing times for the three sites were 17.0 s, 7.9 s, and 6.7 s respectively. Therefore, Site 3 required
the shortest processing time despite its larger raster size, i.e. 9,074,739 grid cells compared with 6 million cells (Site
1) and 4 million cells (Site 2). Figure 9 shows the effect of varying filter kernel size on the processing times for the
Site 1 DSM. Although kernel size was found to strongly impact ROTO run times, in practice the kernel required to
remove the vast majority of OTOs in DSMs is relatively small and usually less than 100. Overall, the ROTO algorithm
was found to be efficient when applied to large rasters. For example, a test in which ROTO was applied with a 25-cell
kernel size to a 289,000,000-grid-cell DSM (the more extensive Rondeau LiDAR data set) required 193.1 s to process.
Figure 3: Site 2 LiDAR intensity image (A) and last-return LiDAR DEM (B).
Figure 4: Site 3 LiDAR intensity image (A) and last-return LiDAR DEM (B).
Figure 5: Site 1 ROTO-derived bare-earth DEM (A) and off-terrain object height (B).
Figure 6: Site 2 ROTO-derived bare-earth DEM (A) and off-terrain object height (B).
Figure 7: Site 3 ROTO-derived bare-earth DEM (A) and off-terrain object height (B).
Figure 8: Comparison of hillshade images for the Black Falls Brook valley bottom (Site 3) derived from the original
DSM (A), ROTO-derived bare-earth DEM (B), and the reference data (C) and for an area within Picton, Ontario (Site 2)
derived from the original DSM (D), ROTO-derived bare-earth DEM (E), and the reference data (F).
Figure 9: The relation between kernel size and processing time for the Site 1 DEM.
6. Discussion
The Kappa analyses demonstrated that the OTOs mapped using the ROTO algorithm (Figures 5B, 6B, and 7B) were
very similar to those that were derived by directly classifying ground points in the raw LiDAR point clouds using a
combination of automated and manual classification procedures. This finding was particularly interesting given the
very different techniques used to derive these two sets bare-earth surfaces and OTO maps. The procedure that was
used to classify the raw LiDAR points was relatively labor-intensive and computationally expensive, relying on LiDAR
point segmentation methods and semi-manual classification. By comparison, the ROTO algorithm maps OTOs solely
based on a raster DSM and is highly automated. This finding is also in contrast to Liu’s (2008) conclusion that OTO
removal methods that are based on processing DSMs suffer from loss of information due to interpolation and lower
accuracy than methods based on processing the raw LiDAR point-clouds directly.
The algorithm requires the user to specify two parameters, namely the slope threshold and the filter kernel size.
The first of these parameters is related to how steep-sided a feature in the WTT surface must be to be considered
an OTO. If this parameter is set too low, some steeper hills may be removed. If, however, the slope threshold is
set inappropriately high, some OTOs may remain in the DEM after processing. Values of this parameter will vary by
application depending on the smoothness of the DSM, which is affected by the interpolation procedure that is used
to create the surface. The use of the WTT surface for analysis, instead of the original elevation surface, meant that
a single slope threshold could be applied to DSMs regardless of variation in topographic slope. Despite the highly
variable topography of the three study sites, experimentation with the slope threshold showed that values in the
range of 12° to 15° were appropriate for application.
The kernel size parameter is used by both the WTT and interpolation subroutines of the ROTO algorithm and the
results showed that processing times were highly dependent on this value (Figure 9). Overall, however, ROTO was
found to be perform efficiently, with processing times for the study site data sets that varied between 6.7 s and 17s for
rasters containing millions of grid cells. Applying Huang et al.’s (1979) efficient filtering method for the WTT reduced
the computational cost of increasing kernel sizes associated with the WTT subroutine. However, the processing time
spent in the interpolation subroutine was more strongly affected by the kernel size parameter. The interpolation
step could be parallelized in future versions of ROTO to further improve algorithm efficiency. Alternatively, ROTO
could be applied iteratively to DSMs containing very extensive OTOs. It is frequently the case that the majority of
OTOs can be removed with a relatively small kernel size. Rather than initially setting the kernel size to a large enough
value to remove the most extensive, and relatively rare, OTO in a data set (often large industrial buildings), it would
be more efficient to process the raw DSM using a smaller kernel size and then apply a second pass of the algorithm
using the larger kernel size. For example, a similar result to the final single-pass Site 1 bare-earth DEM (Figure 5A)
could be achieved using a first-pass kernel size of 11 cells followed by a subsequent second pass with a 75-cell kernel.
The combined processing time of this two-step solution was 11.69 s, 31.2% faster than the single-pass processing
7. Conclusions
This study presented a new method for creating bare-earth DEMs from LiDAR data sets, called ROTO. The new
algorithm has been implemented as open-source software and a corresponding plugin-tool has been created for a
readily available, open-source GIS package. The algorithm is highly automated and has been demonstrated to be
robust under a variety of settings and OTO types. ROTO was applied to three DSMs of sites that demonstrated a
range of land-covers and topography and was compared with reference DEMs of the sites that were created using
ground-point separation based on point-cloud filtering, segmentation, and semi-manual classification methods. The
algorithm was capable of producing satisfactory bare-earth surfaces in all of the tested conditions, which included
agricultural, forested, and urban settings. The resulting bare-earth DEMs were visually very similar to the reference
data, with measured overall accuracies between 0.159 m and 0.499 m at the 95% confidence level. The OTOs mapped
using the new algorithm (based on the DEMs of difference) were found to have high-levels of agreement (Kappa
coefficients ranged from 0.896 to 0.962) with the OTOs that were identified using the reference data set. This work
demonstrates that comparable levels of OTO removal and mapping accuracy can be achieved when the end-user
has access to the interpolated LiDAR raster last-return DSM but not to the point-cloud source data.
8. References
AREFI, H., and HAHN, M., 2005. A morphological reconstruction algorithm for separating off-terrain points from ter-
rain points in laser scanning data. In International Archives of Photogrammetry Remote Sensing and Spatial Information
Sciences, Vol. 36 (3/W19), Enschede, the Netherlands, September 12-14, 2005, pp. 120–125.
KANG-TSUNG, C., and BOR-WEN, T., 1991, The effect of DEM resolution on slope and aspect mapping. Cartography
and Geographic Information Science,18, pp. 69–77.
KIENZLE, S., 2004, The effect of DEM raster resolution on first order, second order and compound terrain derivatives.
Transactions in GIS,8, pp. 83–111.
CHEN, Q., BALDOCCHI, D., GONG, P., and KELLY, M., 2006. Isolating individual trees in a Savanna woodland using
small footprint LiDAR data. Photogrammetric Engineering and Remote Sensing,72, pp. 923–932.
PRINCE, S., 2002. Estimation of tropical forest structural characteristics using large-footprint lidar. Remote Sensing of
Environment,79, pp. 305–319.
Evans, J.S. and Hudak, A.T., 2007. A multiscale curvature algorithm for classifying discrete return LiDAR in forested
environments. IEEE Transactions on Geoscience and Remote Sensing,45(4), pp.1029-1038.
FLOOD, M., 2001. LiDAR activities and research priorities in the commercial sector. International Archives of Pho-
togrammetry and Remote Sensing, XXXIV-3/W4 Annapolis, MD, pp. 678–684.
GUAN, X., and Wu, H., 2010, Leveraging the power of multi-core platforms for large-scale geospatial data processing:
Exemplified by generating DEM from massive LiDAR point clouds. Computers & Geosciences,36, pp. 1276–1282
HODGSON, M.E., JENSEN, J.R., SCHMIDT, L., SCHILL, S., and DAVIS, B., 2003, An evaluation of LIDAR- and IFSAR-derived
digital elevation models in leaf-on conditions with USGS Level 1 and level 2 DEMs. Remote Sensing of Environment,84,
pp 295-308.
Huang, T.S., Yang, G.J. and Tang, G.Y., 1979. A fast two-dimensional median filtering algorithm. IEEE Transactions on
Acoustics, Speech, and Signal Processing, 27(1), pp.13-18.
KOBLER, A., PFEIFER, N., OGRINC, P., TODOROVSKI, L., OŠTIR, K., and DŽEROSKI, S., 2007, Repetitive interpolation: A
robust algorithm for DTM generation from Aerial Laser Scanner Data in forested terrain. Remote Sensing of Environ-
ment,108, pp. 9–23.
LIM, K., TREITZ, P., WULDER, M., ST-ONGE, B., and FLOOD, M., 2003. LiDAR remote sensing of forest structure.
Progress in Physical Geography,27, pp. 88–106.
Li, Y., Yong, B., Wu, H., An, R. and Xu, H., 2014. An improved top-hat filter with sloped brim for extracting ground
points from airborne lidar point clouds. Remote Sensing, 6(12), pp.12885-12908.
LIU, X., 2008, Airborne LiDAR for DEM generation: some critical issues. Progress in Physical Geography,32, 31-49.
LIU, J.G., and MASON, P.J., 2009, Essential Image Processing and GIS for Remote Sensing, 443 p. (Chichester: Wiley-
LINDSAY J.B., 2016. Whitebox GAT: A case study in geomorphometric analysis. Computers & Geosciences,95: 75-84.
DOI: 10.1016/j.cageo.2016.07.003
MENG, X., CURRIT, N., and ZHAO, K., 2010, Ground filtering algorithms for airborne LiDAR data: A review of critical
issues. Remote Sensing,2, pp. 833–860, doi:10.3390/rs2030833.
NELSON, A., REUTER, H.I., and GESSLER, P., 2009, DEM Production Methods and Sources. In Geomorphometry: Con-
cepts, Software, Applications, T. Hengl and H.I. Reuter (Eds.), Series: Developments in Soil Science. pp. 65–85 (Amster-
dam: Elsevier).
POPESCU, S.C., WYNNE, R.H., and NELSON, R.F., 2003. Measuring individual tree crown diameter with lidar and
assessing its influence on estimating forest volume and biomass. Canadian Journal of Remote Sensing,29, pp. 564–
ROTTENSTEINER, F., and BRIESE, C., 2002, A new method for building extraction in urban areas from high-resolution
LIDAR data. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, 34(3/A), 295–
SITHOLE, G., and VOSSELMAN, G., G., 2001, Filtering of laser altimetry data using a slope adaptive filter. International
Archives of Photogrammetry, Remote Sensing, and Spatial Information Sciences, 34(3/W4), pp. 203–210.
Sizintsev, M., Derpanis, K.G. and Hogue, A., 2008. Histogram-based search: A comparative study. In Computer Vision
and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on (pp. 1-8). IEEE.
SOHN, G., and DOWMAN, I.J., 2008, A model-based approach for reconstructing a terrain surface from airborne
LIDAR data. The Photogrammetric Record,23, 170–193.
SOHN, G., and DOWMAN, I.J., 2001, Extraction of buildings from high-resolution satellite data and LIDAR. In The
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences XXXVII, Part B3b, 2008,
Beijing, pp. 693-698.
SUÁREZ, J.C., ONTIVEROS, C., SMITH, S., and SNAPE, S., 2005. Use of airborne LiDAR and aerial photography in the
estimation of individual tree heights in forestry. Computers & Geosciences,31, pp. 253–262.
Tarsha-Kurdi, F., Landes, T., Grussenmeyer, P., and Smigiel, E., 2006. New approach for automatic detection of
buildings in airborne laser scanner data using first echo only. In Proceedings of the ISPRS Commission III Symposium:
Photogrammetric Computer
Vision, PCV’06, 20–22 September, Bonn, Germany, International Archives of Photogrammetry, Remote Sensing and Spatial
Information Sciences, Vol. XXXVI, Part 3, pp. 25–30.
ZHANG, K., CHEN, S.-C., WHITMAN, D., SHYU, M.-L., YAN, J., and ZHANG, C., 2003, A progressive morphological filter
for removing nonground measurements from airborne LIDAR data. IEEE Transactions on Geoscience and Remote
Sensing,41, pp. 872–882.
ZIMBLE, D.A., EVANS, D.L., CARLSON, G.C., PARKER, R.C., GRADO, S.C., and GERARD, P.D., 2003. Characterizing vertical
forest structure using small-footprint airborne LiDAR. Remote Sensing of Environment,87, pp. 171–182.
... filter developed by John Lindsay form University of Guelph (Lindsay J., 2018) was used, followed by a Gaussian filter. ROTO is described as a tool that "is typically applied to LIDAR DEMs which frequently contain numerous off-terrain objects (OTOs) such as buildings, trees and other vegetation, cars, fences and other anthropogenic objects" (Lindsay, 2018). The filter was applied with the WhiteboxTools Python package (Copyright (c) 2017-2019 John Lindsay), with the filter parameter set at 15 and slope, at 10. Gaussian filter was also applied via WhiteboxTools with sigma value set to 1, which corresponds to a 9-pixel kernel size. ...
... Minimum-bin gridding interpolation was used where bathymetry datasets were denser, while Kriging and Natural Neighbour interpolation were preferred in areas covered by sparser datasets. Final DEMs were then filtered with RemoveOffTerrainObjects (ROTO) (Lindsay, 2018) and Gaussian filters. ...
Full-text available
Great Lakes coastal wetlands are strongly influenced by their physical environment, particularly by water-level fluctuations that incorporate the effect of large-scale synoptic systems controlling climate (i.e. surface air temperature, precipitation, evaporation, winds, etc). Water levels are therefore frequently described as the primary driver of changes of wetland structure, composition, and distribution, with annual, seasonal, and sub-seasonal fluctuations maintaining wetland vegetation biodiversity and habitat extent. Wetland ecosystems are hence highly vulnerable to the anticipated global warming, as altered lake hydrology can have a significant impact on their ecology. Adverse effects of climate change, which may result in habitat and/or biodiversity loss or increased anthropogenic stressors and expansion of invasive plant species, may lead to a dramatic deterioration of their conservation and functional values. Great Lakes coastal wetlands are strongly influenced by their physical environment, particularly by water-level fluctuations that incorporate the effect of large-scale synoptic systems controlling climate (i.e. surface air temperature, precipitation, evaporation, winds, etc). Water levels are therefore frequently described as the primary driver of changes of wetland structure, composition, and distribution, with annual, seasonal, and sub-seasonal fluctuations maintaining wetland vegetation biodiversity and habitat extent. Wetland ecosystems are hence highly vulnerable to the anticipated global warming, as altered lake hydrology can have a significant impact on their ecology. Adverse effects of climate change, which may result in habitat and/or biodiversity loss or increased anthropogenic stressors and expansion of invasive plant species, may lead to a dramatic deterioration of their conservation and functional values.
Full-text available
Airborne light detection and ranging (lidar) has become a powerful support for acquiring geospatial data in numerous geospatial applications and analyses. However, the process of extracting ground points accurately and effectively from raw point clouds remains a big challenge. This study presents an improved top-hat filter with a sloped brim to enhance the robustness of ground point extraction for complex objects and terrains. The top-hat transformation is executed and the elevation change intensity of the transitions between the obtained top-hats and outer brims is inspected to suppress the omission error caused by protruding terrain features. Finally, the nonground objects of complex structures, such as multilayer buildings, are identified by the brim filter that is extended outward. The performance of the proposed filter in various environments is evaluated using diverse datasets with difficult cases. The comparison of the proposed filter with the commercial software Terrasolid TerraScan and other popular filtering algorithms demonstrates the applicability and effectiveness of this filter. Experimental results show that the proposed filter has great promise in terms of its application in various types of landscapes. Abrupt terrain features with dramatic elevation changes are well preserved, and diverse objects with complicated shapes are effectively removed. This filter has minimal omission and commission error oscillation for different test areas and thus demonstrates a stable and reliable performance in diverse landscapes. In addition, the proposed algorithm has high computational efficiency because of its simple and efficient data structure and implementation.
Full-text available
[1] Terrain attributes based on upslope contributing area, A, are used widely in distributed hydrologic models. Several grid-based algorithms are available for estimating A. In this study, five algorithms (D8, ρ8, MFD, DEMON, and D∞) were compared quantitatively on two undulating agricultural fields (63 and 109 ha) in northeastern Colorado. Global positioning system (GPS) data (0.02-m accuracy) were used to generate grid digital elevation models (DEMs) at 5-, 10-, and 30-m cell sizes. Relative differences between A values estimated using single- and multiple-direction algorithms increased with decreasing grid cell size. Relative differences were greatest along ridges and side slopes, and differences decreased where the terrain became more convergent. Multiple-direction algorithms (MFD, DEMON, and D∞), allowing for flow divergence, are recommended on these undulating terrains for 5- and 10-m grids where A is most sensitive to the algorithm selection.
This paper describes an open-source geographical information system (GIS) called Whitebox Geospatial Analysis Tools (Whitebox GAT). Whitebox GAT was designed to provide a platform for the rapid development and testing of experimental geospatial analysis methods, supported by its extensible design, integrated facilities for custom plug-in tool authoring, and its novel open-access design philosophy. One of the unique characteristics of Whitebox GAT is the ease with which users can inspect and modify the algorithms for individual geoprocessing tools. The open-access software model that Whitebox GAT adopts is designed to lessen the barriers that are often imposed on end-users when attempting to gain deeper understanding of how a specific function operates. While Whitebox GAT has an extensive range of GIS and remote sensing analytical capabilities, making it broadly suited for advanced scientific research applications in the Earth Sciences, this paper focusses on the software's application in the field of geomorphometry. An airborne LiDAR data set for a small headwater catchment of the Missisquoi River in northern Vermont, USA, was filtered to identify ground-points and then interpolated into a 2.0 m resolution bare-Earth DEM. The DEM was processed to remove spurious off-ground objects (mainly buildings), to reduce surface roughness under heavy forest cover, and to hydrologically pre-condition the DEM. These data were then used to extract salient hydrological structures, i.e. the stream network and their associated sub-basins.
Stream burning is a common flow enforcement technique used to correct surface drainage patterns derived from digital elevation models (DEM). The technique involves adjusting the elevations of grid cells that are coincident with the features of a vector hydrography layer. This paper focuses on the problematic issues with common stream burning practices, particularly the topological errors resulting from the mismatched scales of the hydrography and DEM data sets. A novel alternative stream burning method is described and tested using five DEMs of varying resolutions (1 to 30 arc-seconds) for an extensive area of southwestern Ontario, Canada. This TopologicalBreachBurn method uses total upstream channel length (TUCL) to prune the vector hydrography layer to a level of detail that matches the raster DEM grid resolution. Network pruning reduces the occurrence of erroneous stream piracy caused by the rasterization of multiple stream links to the same DEM grid cell. The algorithm also restricts flow within individual stream reaches, further reducing erroneous stream piracy. In situations where two vector stream features occupy the same grid cell, the new tool ensures that the larger stream, designated by higher TUCL, is given priority. TUCL-based priority minimizes the impact of the topological errors that occur during the stream rasterization process on modeled regional drainage patterns. The test data demonstrated that TopologicalBreachBurn produces highly accurate and scale-insensitive drainage patterns and watershed boundaries. The drainage divides of four large watersheds within the study region that were delineated from the TopologicalBreachBurn-processed DEMs were found to be highly accurate when compared with the official watershed boundaries, even at the coarsest grid resolutions, with Kappa index of agreement values ranging from 0.952 to 0.921. The corresponding Kappa coefficient values for a traditional stream burning method (FillBurn) ranged from 0.953 to 0.490, demonstrating a significant decrease in mapping accuracy at coarser DEM grid resolutions.
Digital elevation models (DEMs) that are used in hydrological applications must be processed to remove sinks, mainly topographic depressions. Flow enforcement techniques include filling methods, which raise elevations within depressions, breaching, which carves channels through blockages, and hybrid methods. Despite previous research demonstrating the large impact to DEMs and subsequent analyses of depression filling, it is common practice apply this technique to flow enforcement. This is partly due to the greater efficiency of depression filling tools compared to breaching counterparts, which often limits breaching to applications of small- to moderate-sized DEMs. A new hybrid flow enforcement algorithm is presented in this study. The method can be run in complete breaching, selective breaching (either breached or filled), or constrained breaching (partial breaching) modes, allowing for greater flexibility in how practitioners enforce continuous flow paths. Algorithm performance was tested with DEMs of varying topography, spatial extents, and resolution. The sites included three moderate sized DEMs (52,000,000 to 190,000,000 cells) and three massive DEMs of the Iberian Peninsula, and the Amazon and Nile River basins, the largest containing nearly one billion cells. In complete breaching mode, the new algorithm required 87% of the time needed by a filling method to process the test DEMs, while the selective breaching and constrained breaching modes, operating with maximum breach depth constraints, increased run times by 8% and 27% respectively. Therefore, the new algorithm offers comparable performance to filling and the ability to process massive topographic data sets, while giving practitioners greater flexibility and lowering DEM impact.
A new procedure for the representation of flow directions and calculation of upslope areas using rectangular grid digital elevation models is presented. The procedure is based on representing flow direction as a single angle taken as the steepest downward slope on the eight triangular facets centered at each grid point. Upslope area is then calculated by proportioning flow between two downslope pixels according to how close this flow direction is to the direct angle to the downslope pixel. This procedure offers improvements over prior procedures that have restricted flow to eight possible directions (introducing grid bias) or proportioned flow according to slope (introducing unrealistic dispersion). The new procedure is more robust than prior procedures based on fitting local planes while retaining a simple grid based structure. Detailed algorithms are presented and results are demonstrated through test examples and application to digital elevation data sets.
Essential Image Processing and GIS for Remote Sensing is an accessible overview of the subject and successfully draws together these three key areas in a balanced and comprehensive manner. The book provides an overview of essential techniques and a selection of key case studies in a variety of application areas. Key concepts and ideas are introduced in a clear and logical manner and described through the provision of numerous relevant conceptual illustrations. Mathematical detail is kept to a minimum and only referred to where necessary for ease of understanding. Such concepts are explained through common sense terms rather than in rigorous mathematical detail when explaining image processing and GIS techniques, to enable students to grasp the essentials of a notoriously challenging subject area. The book is clearly divided into three parts, with the first part introducing essential image processing techniques for remote sensing. The second part looks at GIS and begins with an overview of the concepts, structures and mechanisms by which GIS operates. Finally the third part introduces Remote Sensing Applications. Throughout the book the relationships between GIS, Image Processing and Remote Sensing are clearly identified to ensure that students are able to apply the various techniques that have been covered appropriately. The latter chapters use numerous relevant case studies to illustrate various remote sensing, image processing and GIS applications in practice.
Article The detailed topographic information contained in light detection and ranging (LiDAR) digital elevation models (DEMs) can present significant challenges for modelling surface drainage patterns. These data frequently represent anthropogenic infrastructure, such as road embankments and drainage ditches. While LiDAR DEMs can improve estimates of catchment boundaries and surface flow paths, modelling efforts are often confounded difficulties associated with incomplete representation of infrastructure. The inability of DEMs to represent embankment underpasses (e.g. bridges, culverts) and the problems with existing automated techniques for dealing with these problems can lead to unsatisfactory results. This is often dealt with by manually modifying LiDAR DEMs to incorporate the effects of embankment underpasses. This paper presents a new DEM pre-processing algorithm for removing the artefact dams created by infrastructure in sites of embankment underpasses as well as enforcing flow along drainage ditches. The application of the new algorithm to a large LiDAR DEM of a site in Southwestern Ontario, Canada, demonstrated that the least-cost breaching method used by the algorithm could reliably enforce drainage pathways while minimizing the impact to the original DEM.