The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1^
2.2 Remote Sensing Payload
A market evaluation at the time of the study showed that there
were no suitable multispectral sensors available which could
have provided multi-channel imagery in the bands Red, NIR
(and preferably also Green) at a total system weight of 1 kg or
less. In order to proceed with the initial feasibility tests in the
2006 vegetation period, it was decided to use a combination of
two different sensors for RGB and NIR, which had to be flown
in two subsequent flight missions. The two imaging sensors
were:
NIR sensor Sony SmartCam - 1/2" monochrome CCD with
1280x1024 pixels; onboard CPU with 400 MHz CPU, 256 MB
DDR-SDRAM, and Windows XP embedded; Compact Flash
on-board storage up to 4 GB; camera body weight 400g (w/o
optics or power supply); precision optical lenses from
Schneider Kreuznach, corrected from 400 to 1000 nm; high-
pass filter with a cut-on wavelength of 780 nm (to approx. 1000
nm).
Figure 2: Mini UAV of weControl (Zurich)
with the RGB sensor Canon EOS 20D.
RGB sensor Canon EOS 20D - commercial digital SLR camera;
CMOS chip with 22.5 x 15.0 mm and 3504 x 2336 pixels;
weight 770 g; radiometric resolution (RAW format) of 12 bits;
standard lens with fixed focal length; IR blocking filter with
cut-off wavelength at approx. 720 nm.
The optical lenses of the two camera systems were chosen so
that they would yield the same ground sampling distances
(GSDs) from identical flying heights.
2.3 Field Test Campaign
Due to the available sensor constellation, the acquisition of the
RGB and NIR imagery with a GSD of approx. 7 cm had to be
carried out in two separate pre-programmed flights. The flights
were planned as a photogrammetric strip with 4 images and an
overlap of approx. 60%. The imagery was acquired at a flying
height of 100 metres above ground resulting in image scales of
approx. Ull'OOO (for the RGB imagery) and 1:13750 (for the
NIR imagery). Precise ground control with an accuracy of
approx. 2 cm horizontal and 3 cm vertical was established
(Annen u. a., 2007).
Figure 3: Section of the generated false colour true orthoimage.
2.4 Processing and Results
The goal of the processing phase was to derive plant health
figures for each plot which can directly be related to those used
by the field specialists, who normally express the plant status in
percentage of damaged leaves. The processing steps can be
summarised as follows (Brosi, 2006) and (Annen u. a., 2007):
• extraction of raw imagery from both sensors
• geo-referencing of Red and NIR imagery
• co-registration of Red & NIR imagery by a true
orthoimage
• radiometric corrections using radiometric field targets
• masking out of 'background' soil using a GIS-based
geometric buffer for the grapevine rows
• calculation of uncalibrated or 'raw' NDVI values
• matching of raw NDVI values with leaf damage values of
control plots by means of a weighted linear regression
• calculation of leaf damage values for all test plots by
means of linear interpolation
The described sensor and flight constellation, namely the use of
two different and radiometrically uncalibrated 'off-the-shelf
sensors and the spatially disparate exposure centres for the RGB
and NIR imagery, introduced a number of challenges into the
processing chain. The very special challenge of co-registering
the very high-resolution RGB and NIR channels acquired in
separate flights to an accuracy of 1-2 pixels at a GSD of only 7
cm was successfully met by generating a true orthoimage (see
Figure 3). The underlying elevation data modelling the surface
of the grapevines was derived using the GIS-based position
information of the grapevine rows in combination with an
average height for the grapevine cultivation.
In the subsequent investigations a number of vegetation indices
were successfully used to assess the plant health within a grape
vine test field. However, it is shown in (Brosi, 2006) that
excellent and robust results can be achieved by using the
standard NDVI. In the case our grapevine test field the
percentage of damaged leaves determined with remote sensing
agreed to within 10% with the comprehensive ground truth
information (see Figure 1). This indicates that the remotely
sensed solution roughly yields the same accuracy level as the
very labour intensive ground-based bonification.
1196