rious river
anagement
se of laser
modelling.
ial of laser
bilities for
try for the
ally, laser
elengths.
s for soil
tion works
anagement
parts of the
; the river
n to verify
r altimetry
1), (Cobby
agraphs the
nore detail.
'acy will be
ge capacity
tance (flow
roughness
actors like
ial patterns
. The main
forecasting
nation with
e paragraph
ping of the
gularly.
ie required
jlains. One
'ss informa-
vestigations
of the data
t concluded
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
that it is difficult to distinguish lower types of vegetation
including grass lands, bare arable land and herbaceous
vegetation. However, there are other studies which show that
vegetation height, which can be derived from laser altimetry
data, can be used as a surrogate for the product of stem density
and flexural rigidity (Mason, 2003) and thus would be
sufficient for the water level forecasting model which they use.
In the Netherlands, it is common practice to derive the
roughness parameters from vegetation structure classes.
Rijkswaterstaat employs a list with the hydraulic resistance
values for about 30 different vegetation structure types which
normally appear in Dutch riparian forelands (Velzen et al.,
2003). Up to now, these vegetation structure classes are
manually mapped from aerial stereo-phothographs. This is a
time consuming and expensive method. Therefore, we are
investigating the possibility of automatically or semi-
automatically deriving the vegetation structure classes from
remote sensing images such as airborne hyperspectral data in
combination with laser altimetry data. For this purpose we
investigated the benefit of fusing hyper-spectral data from the
airborne line-scanner CASI (10 bands ranging from 440 nm up
to 870 nm, pixel size 2 x 2 m?), acquired in summer 2003, with
very dense laser altimetry data (30-40 points per m?). The laser
data was acquired in March of 2003 with the helicopter borne
FLI-MAP system from Fugro Inpark (see also paragraph 4).
The test area was the floodplain Gameren along the river Waal
which comprises about 160 ha.
Before land use classification a quite labour-intensive geometric
correction of the CASI data was necessary which has been done
with the ‘rubbersheeting’ algorithm of ERDAS Imagine. To get
the vegetation structure classes from the spectral and laser data
a classification was performed using the software eCognition
(see fig. 1). It follows the concept that important semantic
information is not represented in single pixels but in meaningful
image objects and their mutual relations. Therefore, the image
classification is based on image segments rather than individual
pixels. In a first step, the software extracts homogeneous image
segments in any chosen resolution which are subsequently
classified by means of fuzzy logic.
As input layers we chose all 10 CASI bands, a NDVI-layer
(NDVI = Normalized Difference Vegetation Index) which was
computed from the red and nearby infrared band, the unfiltered
laser heights and a band comprising maximal height differences
input layers
classified image
for hydraulic
roughness
as
spectral
data
Y
laser CTI
data E classification
with
eCognition
Figure 1. Principal of vegetation classification with eCognition
based on spectral and laser altimetry data.
of laser points within one pixel indicating the vegetation height.
All data were resampled to 1.5 x 1.5 m^ pixel size.
After segmentation into small segments, a top down tree
structure approach (hierarchical network) for classification was
pursued. First, two main classes (“land” and “‘water”) were
distinguished. These two classes were further subdivided into
more detailed classes, c.g. the class “land” was subdivided into
the classes “vegetation”, “bare soil” and “shadow” and the class
"water" into "water-plants/duckweed" and “water without
plants”. Totally, 15 classes have been distinguished. The
classification was done by a combination of a standard
supervised Maximum Likelihood classification and rules
(constraints) translated into membership functions. Ecognition
also allows using shape parameters of the segments for
classification, but up to now we did not use this option. The
overall classification accuracy unfortunately has not vet been
quantified, but is estimated to be around 70-80%. A thorough
evaluation of the achieved results still has to be done. However,
it is already apperent that the combination of spectral and height
data is very promising, especially for distinguishing the lower
vegetation structure classes which seemed to be difficult with
laser altimetry alone (Asselman, 2002).
3. VOLUMES
For construction works in the floodplains, such as dyke
displacements and lowering parts of the floodplains in order to
give the river more space for discharging especially with high-
water conditions, detailed and precise DEM’s are required for
soil volume determination. The volume of the soil which have
to be moved or digged up is an essential parameter for the
contracts with the construction firms. In addition, DEM’s are
measured after termination of the work (so called ‘end
models’). Currently these DEM’s are measured with tachymetry
or GPS with about 140 points per ha. The question was whether
laser altimetry could be an alternative with regard to quality and
costs.
Volumes can be calculated by multiplying the mean height with
the concerned area: volume = length x width x height. see fig.
2. Thus the precision of the volume is closely related to the
precision of the mean height of an area. Therefore. the effect of
for example laser scanner point noise can be neglected. The
volume precision also depends on how much soil has to be
digged up or moved, the so-called digging depth (h). A
maximal volume error of 5% is requested. This yields for the
precision of the mean height:
204 <(h/100)-5 or 6, < (h/100)-(5/2).
Tabel | shows the required precision for the mean height as
function of the digging depth.
1 2 3 4
GO, [em] < 2.5 5 7.5 10
Tabel 1. Required precision of the mean height depending on
the digging depth.
digging depth [m]
The determination of the mean height precision of an area
acquired with laser altimetry is based on error propagation of
four error components with different amplitudes and spatial
resolutions. This is described in detail in (Crombaghs et al.,
2002). With increasing area size, the precision of the mean
height decreases because then more error components are
averaged. Our computations showed that the previously
mentioned FLI-MAP system meets the demand of 2,5 cm height
|
|
|
|
|
|
|
