Content uploaded by Roy A. Armstrong
Author content
All content in this area was uploaded by Roy A. Armstrong
Content may be subject to copyright.
NASA Technical Memorandum 112218
1
Grapevine Remote sensing Analysis of Phylloxera Early Stress (GRAPES):
Remote Sensing Analysis Summary∗
Brad Lobitz1, Lee Johnson2, Chris Hlavka3, Roy Armstrong4, and Cindy Bell5
1Johnson Controls World Services, Inc., Ames Operation (currently California State
University, Monterey Bay)
2Johnson Controls World Services, Inc., Ames Operation (currently California State
University, Monterey Bay)
3NASA Ames Research Center, Earth Science Division
4Johnson Controls World Services, Inc., Ames (currently Univ. Puerto Rico)
5Johnson Controls World Services, Inc., Ames Operation (currently California State
University, Monterey Bay)
Abstract
This document describes image processing analysis applied to high spatial resolution
airborne imagery acquired in California's Napa Valley in 1993 and 1994 as part of the Grapevine
Remote sensing Analysis of Phylloxera Early Stress (GRAPES) project. Investigators from
NASA, the University of California, the California State University, and Robert Mondavi
Winery examined the application of airborne digital imaging technology to vineyard management,
with emphasis on detecting the phylloxera infestation in California vineyards. Phylloxera
infestation is a significant problem because the root louse causes vine stress that leads to
grapevine death in three to five years. Eventually the infested areas must be replanted with
resistant rootstock. Visual symptoms of phylloxera infestation include leaf chlorosis, vine size
reduction, and collapse of fruit tissue during the growing season. Increased leaf temperatures have
also been hypothesized for affected vines. Early detection of infestation and changing cultural
∗ full citation: Lobitz, B., L. Johnson, R. Armstrong, C. Hlavka and C. Bell. 1997.
Grapevine Remote Sensing Analysis of Phylloxera Early Stress (GRAPES):
Remote Sensing Analysis Summary. NASA Technical Memorandum No. 112218,
December 1997.
NASA Technical Memorandum 112218
2
practices to compensate for vine damage can minimize crop financial losses from damage and
replanting. Vineyard managers need improved information to decide where and when to replant
fields or sections of fields.
Multi-year airborne images of a vineyard were spatially co-registered so annual relative
changes in leaf area due to phylloxera infestation could be located. Image processing analysis was
applied to data from the Compact Airborne Spectrographic Imager (CASI, imagery acquired in
1993) and the Electro-Optic Camera (EO Camera, imagery acquired in 1994). Changes were
determined by using information obtained from computing Normalized Difference Vegetation
Index (NDVI) images. As the canopy leaf area of infested regions decreased, these regions became
increasingly non-uniform. Infestation spread was also projected in advance using proximity
analysis, a geographic information system (GIS) technique. Two other methods of monitoring
vineyards through imagery were also investigated: optical sensing of the Red Edge Inflection
Point (REIP), and thermal sensing. These did not convey the stress patterns as well as the NDVI
imagery and require specialized sensor configurations. NDVI-derived products are recommended
for monitoring phylloxera infestations.
Introduction
Phylloxera (Daktulosphaira vitifoliae Fitch) affects a number of the grape growing
counties in California and is currently a severe problem in Napa and Sonoma Counties. The
parasitic action of this root louse causes leaf chlorosis, decreases shoot and leaf growth and fruit
yield, and leads to vine death three to five years from onset. Once established, the infestation
spreads quickly through a vineyard. The grapevines' deep rooting pattern makes pesticides
ineffective and there is no known biological control (Granett et al., 1987 and 1991).
Changing cultural practices, such as adjusting pruning severity or changing the amount of
NASA Technical Memorandum 112218
3
irrigation, are sometimes used to prolong fruit production, but there is little the grape grower can
do to combat phylloxera, except replant on phylloxera-resistant rootstock. As the infestation
spreads, the grape yield decreases each year, until the yield is too low justify the maintenance
costs and the vines are plowed up. Replanting is expensive, and the field will also be out of
production for a number of years before the newly planted vines bear fruit.
The Grapevine Remote sensing Analysis of Phylloxera Early Stress (GRAPES) project
was a collaboration between NASA Ames Research Center, the University of California Davis,
the University of California Cooperative Extension, California State University Chico, and
Robert Mondavi Winery (Oakville, CA). The project was developed to demonstrate the use of
remotely sensed data for vineyard management, with emphasis on monitoring phylloxera
infestation, e.g., using remotely sensed data to help decide when to replant a phylloxera infested
field.
Some vineyard managers have used aerial photography to study phylloxera spread
(Wildman, 1983). The GRAPES project incorporated airborne digital imaging systems with
subsequent image processing and analysis to enhance information content with respect to canopy
size (Johnson et al., 1996). This report summarizes methodology used to generate annual imagery
for vineyard managers to monitor the spread of phylloxera in a Mondavi vineyard. This
methodology could be applied to digitally acquired imagery and to film-based traditional aerial
photography that has been scanned into digital form. Satellite images acquired by the Landsat
Thematic Mapper (TM) and Satellite Pour l'Observation de la Terre (SPOT) were also
purchased, but were used for valley wide analysis and not at the vineyard or block scale. (A
block is the smallest management unit within a vineyard.) This document also describes the image
processing analysis applied to high spatial resolution airborne imagery acquired in California's
Napa Valley in 1993 and 1994. Results of the analysis indicate the procedures used offer tangible
NASA Technical Memorandum 112218
4
benefits to growers.
Study Site and Ground Data
Airborne digital data were acquired over the study site, ToKalon Ranch (a vineyard
owned and managed by the Robert Mondavi Winery), in 1993 and most of Napa Valley,
including ToKalon and two other Mondavi vineyards: Carneros and Oak Knoll, in 1994.
ToKalon ranch lies on the West edge of the valley floor and is surrounded on the other three sides
by other vineyards. The ToKalon soils are mainly Bale loam and clay loam with 0-2% slope,
with some Bale clay loam with 2-5% slope and some Coombs gravelly loam with 0-2% slope.
The southern portion of the ranch is Clear Lake clay (2-5% slope). Images from 1993 and 1994
covering all of ToKalon Ranch were processed, but this report focuses on one five-hectare block
(denoted as block I in the following discussion) within ToKalon. Block I consisted of cabernet
sauvignon grapevines on AXR-1 rootstock with four-meter row spacing planted primarily in
Clear Lake clay soil. The analysis of this block was used to illustrate the type of information that
can be generated for an entire vineyard.
Nine plots within block I were used for collecting field data (pruning weights, phylloxera
counts, and leaf samples). These plots were chosen based on 1992 aerial color infrared
photography and a pre-growing season phylloxera study. Interpretation of the color infrared
photos provided locations of infested areas. Infestation levels were then confirmed by root
digging in the field. The nine plots consisted of three plots each of uninfested, mildly infested,
and severely infested vines. Each plot contained forty vines. At the end of each growing season
(January), Mondavi pruned the vines and weighed vegetative material. Infrared temperatures
were also collected in the field, using a hand held infrared thermometer, from five representative
plots, soil, and roads in 1994. These data were used to support airborne thermal infrared data
NASA Technical Memorandum 112218
5
collection.
Aircraft Data
Aerial photography (acquired by Ames Research Center's C-130 and ER2 aircraft) and
digital imagery from a variety of airborne sensors were used during the GRAPES project. Several
airborne sensors with different specifications and spectral characteristics were used to investigate
block monitoring capabilities and to test the utility of the digital image processing methods.
The airborne sensors used included: Airborne Data Acquisition and Registration (ADAR),
Airborne Infrared Disaster Assessment System (AIRDAS), Compact Airborne Spectrographic
Imager (CASI), Digital Multi-Spectral Video (DMSV), Electro-Optic Camera (EO Camera),
NS001 Thematic Mapper Simulator (TMS), and Real Time Digital Airborne Camera System
(RDACS). Each of these sensors can be used to acquire image data at different spatial, spectral,
and radiometric resolutions. Spatial resolution, or ground resolution, is the size of the smallest
area element that can be detected for the image. Spatial resolution, referred to as pixel size,
depends on the sensor platform (aircraft) collection altitude. For example, the CASI sensor can be
used to acquire image data at spatial resolutions between 0.6 m and 10 m. At 1200 m altitude, the
CASI system yields a pixel size of 1.6 m (rounded to 2 m in this report) and at 450 m, 0.6 m.
The total area imaged, therefore, decreases with decreasing aircraft altitude. Below some limiting
height the sensor systems cannot acquire and store data fast enough to provide continuous
ground coverage. The AIRDAS (Ames Research Center) sensor, which was designed for fire
monitoring, has two thermal infrared channels (Ambrosia et al., 1994). The AIRDAS low
temperature, thermal channel (9250 nm center wavelength) was the most interesting for the
GRAPES project, because it was used to determine surface (brightness) temperatures. The CASI
can be used to collect data in spatial mode and spectral mode. In the CASI's spatial mode the
NASA Technical Memorandum 112218
6
sensor functions as a push-broom imager with up to 15 bands and in spectral mode the sensor
operates like a group of spectrometers (1.8 nm spectral resolution) sweeping the flight path
(Borstad Assoc., 1991). Only the spatial mode with four or eight channels and a resampled
spatial resolution of about 2.0 m were used for the project. The four channel configuration was
used to provide false color infrared imagery and the eight channel configuration provided data
along the red edge of the vegetation reflectance curve, Figure 1. The DMSV had four similar
channels (Lyon, 1994). The EO Camera, flown at an altitude of 20 km aboard the NASA ER-2,
has a nominal spatial resolution of five meters and was flown with five channels. The NS001
TMS has eight channels, was flown aboard the NASA C-130 with a spatial resolution of three to
five meters. Seven of these channels correspond to the TM instrument channels. The RDACS
has three cameras with narrow band filters (about ten nanometers). While the imagery acquired
from all these sensors was examined, most of the data analysis was performed on data from the
primary project sensors (CASI and EO Camera). Data from the other sensors were provided by
their representatives for evaluation. Only the CASI and EO Camera processing will be discussed
in the following sections. The central wavelengths of each spectral channel, the spatial resolution,
and the sensor provider for these sensors are summarized in Table 1. For more detailed sensor
specifications, see the Appendix.
Table 1. Airborne sensors used in the GRAPES project. Spatial resolutions are based on the
respective flight altitudes.
Sensor Channel centers (nm) Spatial res. (m) Source
ADAR 440, 515, 650, 890 3 Positive Systems, Whitefish
MT
AIRDAS 645, 1635, 4550, 9250 5 NASA Ames Res. Ctr.
CASI (4 ch) 550, 630, 680, 787 2 Borstad Associates, Sidney,
BC, Canada
NASA Technical Memorandum 112218
7
CASI (8 ch) 500, 550, 630, 680, 710, 737,
747, 788 resampled to 2 Borstad Associates, Sidney,
BC, Canada
DMSV 450, 550, 650, 750 1 SpecTerra Systems, Pty.,
Ltd., Nedlands, Western
Australia
EO Camera 680, 720, 735, 750, 775 5 NASA Ames Res. Ctr.
NS001 TMS 489, 566, 665, 839, 1240,
1640, 2240, 11300 3 to 5 NASA Ames Res. Ctr.
RDACS 548, 650, 821 1 NASA Stennis Space Ctr.
The four-channel CASI data were acquired 26 July 1993 using a nadir view over all of
ToKalon at an altitude of 1200 m. An eight-channel, 45 degree oblique view was also acquired
over block I at an altitude 450 m respectively. An oblique image minimizes the spectral effect of
the soil, since the sensor looks at an angle to the canopy. Some soil was still seen because the
vines have an open canopy and the vine rows are widely spaced. The oblique image was used
only for the red edge inflection point calculations. The CASI data were delivered as radiance data,
with no atmospheric corrections applied. The EO Camera data were acquired 1 August 1994 at
an altitude of twenty kilometers. Reliable radiance calibration data were not available for the EO
Camera at the time, so these data were used uncalibrated.
The processing described below was performed with imagery from the study site, but
focusing on block I, since field data were available for ground truth for this block. Since the blocks
were known to be infested with phylloxera based on root diggings and damaged regions had a
rapid annual spread rate, the stress or damage to the vines was concluded to be due to phylloxera.
Preliminary work with a proximity search was also accomplished using four vineyard blocks (C,
H, I, and J) to study the spread of phylloxera within the blocks. From their appearance in the
false color infrared imagery, two of the blocks, H and I, were lightly stressed in 1993, while C
NASA Technical Memorandum 112218
8
and J were moderately stressed in 1993. In 1994, blocks H and I were moderately stressed and
block C was severely stressed and block J was moderately stressed.
A 1993 C-130 1:6 000 scale and a 1994 ER2 1:32 000 scale photograph were also
available for the ToKalon Ranch region. The photographs were scanned to generate multispectral
images, analogous to the NIR, red, and green channels of an airborne scanner. The effective pixel
size for these images after registration was one meter.
Processing and Results
There were several factor affecting the image analysis procedures: (1) the image data had
to be in the same map projection as, and spatially co-registered with, the other data layers used in
the GRAPES project (e.g., soils, road network, hydrology); (2) the imagery was to be compared
from year to year; and (3) the data analysis procedures needed to be practical and applicable for
procedural repeatability and ease of use. The third factor required choosing a procedure providing
results easily comparable with the field data.
The second constraint was the most difficult to achieve. The data processing had to
reconcile differences in sensor characteristics and provide results that were not affected by
differences in viewing conditions. Several normalization schemes were tested to reconcile the data
sets. The scheme finally selected was chosen based on ease of implementation and validity of the
results. The simplest procedure was to match the spatial resolutions of the data, then classify the
imagery based on the Normalized Difference Vegetation Index (NDVI) values. This simple
method provided sufficient results without complicated sensor calibration and atmospheric
correction models applied to the imagery.
Image Registration
NASA Technical Memorandum 112218
9
The four-channel CASI imagery acquired in 1993 was used as the base date, while
subsequent airborne digital images were geo-registered to it. The ToKalon site data, consisting of
three adjacent passes, were mosaicked together into one image. The mosaicked image was
registered to a Universal Transverse Mercator (UTM) zone ten projection, with two-meter
spatial resolution. This image registration was accomplished using global positioning system
(GPS) data points collected in the field. These points were then located in the image and used as
ground control points (GCP's). Finally, the image was warped so the GCP's were in the correct
positions relative to the UTM coordinate system. Later images of the same area, such as the EO
Camera imagery, were registered to this image. A standard image to image transformation
procedure was used. In this case the GCP's were pixel locations in the CASI image and the
equivalent pixel locations in the unregistered image.
Equalization of Spatial Resolution
The EO Camera imagery, acquired in 1994, was registered to the 1993 CASI imagery. The
1994 EO Camera data had a nominal spatial resolution of five meters, while the CASI imagery
had a spatial resolution of two meters. The resolution difference was compounded by the EO
Camera lens could not be focused across all wavelength channels simultaneously, and
consequently the channels, particularly the red channel, were slightly out of focus. The 1994
image was registered to the 1993 image, while adjusting for the pixel size (sampling interval)
difference by resampling. Low pass (averaging) spatial filters of various sizes were used to
degrade the CASI and EO Camera near infrared (NIR) to the EO Camera red spectral band spatial
resolution. The best visual match was obtained with a 5x5 window applied to the EO Camera
NIR band and a 7x7 window applied to both of the CASI channels. This procedure normalized
the EO Camera focusing problems, Figure 2.
NASA Technical Memorandum 112218
10
Normalized Difference Vegetation Index Analysis
The NDVI was next applied to the imagery. The NDVI is defined as
The NDVI highlights differences in vegetation canopy reflectance. Healthy vegetation has strong
absorbance characteristics in the red portion of the electromagnetic spectrum (EMS), while also
reflecting strongly in the NIR portion of the EMS. These properties are due to the interaction of
light with the chlorophyll in the plant tissue, Figure 1. Subtle changes in vegetation vigor or leaf
chlorophyll composition result in subtle alterations in absorbance and reflectance characteristics.
These characteristics are then highlighted in the NDVI. The index is near zero for bare soil, but
can be close to 1.0 for a dense, healthy canopy. The NDVI was used because it lessens the
influence of solar illumination, angular influences, slope, and viewing geometry. It performs
consistently between sensors, for different flights, and within the images. NDVI is also correlated
to leaf area index (LAI), or canopy leaf amount, and biomass (Tucker 1979). The index
compensates for brightness differences and highlights the spectral differences between pixels.
Absolute NDVI's were not directly comparable because of year-to-year differences in non-
canopy variables and non-phylloxera related growth effects. Non-canopy variables include
calibration differences (the CASI data were calibrated to radiance versus the raw EO Camera
data), atmospheric conditions (weather conditions and aerosol concentration), and solar
illumination angle differences. Year to year plant growth differences could be also be a response
to other factors, including other plant stresses, changes in management practices, and increased
rainfall (i.e., more irrigation) in 1994. However, if the range of NDVI values within the images is
represented by classes, then relative values (classes) in the images from the same areas on
NDVI = NIR-red
NIR+red .
NASA Technical Memorandum 112218
11
different dates can be compared. A small number of classes makes the images easier interpret.
Initially NDVI data were assigned hues ranging from brown (bare soil), through yellow (small or
stunted vines), to dark green (vigorous growth). This approach showed damage patterns within
the vineyard. Later, images were coded using a rainbow (color spectrum) color coding due to the
greater hue separation. This scheme was preferred by the Mondavi vineyard manager and
subsequently because the coding of choice for the images.
Subjective comparisons of NDVI images of block I from 1993 and 1994 were difficult, so
an unsupervised classification was used to categorize block I and, later, the entire vineyard. An
objective method of determining class breaks was needed, so Iterative Self-Organizing Data
Analysis (ISODATA, Duda and Hart, 1973), an unsupervised classification algorithm was used.
Utilized with only one input image band, the ISODATA routine determines the clusters within
the range of pixel values in the image using the number of clusters the user inputs. The
ISODATA classification process begins by dividing the range of values and using the midpoint of
each breakpoint as the starting means for the number of classes specified by the user. Each pixel
is then assigned to the cluster that has the closest mean value to the pixel value. Cluster means are
then recomputed based on the pixels assigned to the clusters, and the pixels are again assigned to
clusters based on the new means. Eventually the means settle down and the process terminates.
When run on block I, six classes were used in the classification. This kept the number of classes
down ease of interpretation, while still representing the image variation. For the entire vineyard,
this number was doubled to twelve classes, since there was much variation in NDVI values due to
differences in vine maturity, trellis type, and vine spacing as well as plant condition.
A number of vineyard blocks were pulled after the 1993 data acquisition, and,
subsequently, large areas of bare soil were evident in the 1994 imagery. These "soil" pixels would
be over-represented in the classes generated from the 1994 NDVI image. To equalize the
NASA Technical Memorandum 112218
12
distribution, the NDVI images were visually compared with aerial photos and the false color
infrared imagery to select an NDVI threshold that masked out the nearly bare and bare soil pixels.
Any pixels below the threshold (the soil) became zero and were not considered in the
classification routine. The thresholded image contained only vegetated landscape elements and the
range of NDVI values was therefore reduced. The proportion of high NDVI pixels appeared
stable, since those pixels were primarily trees along the streams, roads, and hillside as well as the
vigorous grapevines.
The ISODATA classification was performed to the filtered CASI (1993) and EO Camera
(1994) images. A common area covering the ToKalon vineyard was used in these steps for both
years. Subsets of each of these classified images for block I are shown in Figure 3. The mean
class values for each of the plots within block I and their pruning weights are shown in Table 2.
A percentage summary of the classification for block I by class values is shown in Figure 4. In
the vineyard as a whole, there were NDVI class values below and above the block I class values.
A few NDVI classes predominated in 1993 (Figure 4). In 1994, the vineyard blocks had a large
number of bare soil pixels and a broad distribution of other class values, because phylloxera
damage lead to a decrease in block uniformity.
The 1994 histogram for block I (Figure 4) also showed increases in high class values,
which was consistent with the pruning weights given for plots 7 - 9 (Table 2). Infestation levels
were evident in the pruning weights and to a lesser extent in the NDVI class values. Table 2
indicates the mean class values for the pixels in infested plots 1 - 6 had decreased, while the mean
class values for the lightly or unaffected plots (7, 8, and 9) had increased. Class values were
plotted against pruning weight in Figure 5. Correlations between the pruning weights and either
NDVI or class values were similar, though greater in 1994 versus 1993: R2 was about 0.60 for
1993, 0.86 for 1994. The correlation between pruning weight and NDVI class for both years
NASA Technical Memorandum 112218
13
combined (R2 = 0.76) was higher than the correlation between all of the pruning weights with all
NDVI values (R2 = 0.66). Therefore, classifying the imagery improved the relationship between
pruning weights and NDVI. Similar improvement occurred with the NDVI values when the red
and NIR channels were radiometrically normalized before computing the NDVI (results not
presented here).
Table 2. Mean pruning weights per vine and mean Normalized Difference Vegetation Index class
values for pixels within plots.
1993 1994
Plot Pruning
Weight (kg) Class Pruning
Weight (kg) Class
1 1.39 1.648 0.92 0.023
2 0.70 1.129 0.45 0.186
3 0.95 2.581 0.58 0.795
4 2.25 5.835 1.57 2.000
5 2.94 6.626 2.50 3.907
6 1.13 4.851 0.54 0.020
7 1.60 7.301 2.45 8.375
8 2.51 6.699 3.36 7.488
9 2.95 6.720 3.95 9.085
R2 = 0.60 R2 = 0.88
A difference image was generated for block I (Figure 6) to spectrally compare the images.
Image differencing was accomplished by subtracting 1994 NDVI class image from the 1993 class
image. An extensive area, in the middle of the vineyard block indicated a significant NDVI
decrease associated with the spread of phylloxera. The dark gray areas on the left side of the
NASA Technical Memorandum 112218
14
vineyard block indicate canopy cover increases. This image also exhibited large decreases centered
on the areas with lower class values in 1993, such as at the upper middle of the block (Figure 3).
NDVI analysis was also performed with the scanned color infrared photographs from
block I. Comparison of the 1993 NDVI classes with the 1993 pruning weights resulted in an R2
of 0.64. Comparisons of the 1994 classes and 1994 weights resulted in an R2 of 0.90. NDVI
values from the 1993 scanned photograph compared to the 1993 pruning weights resulted in an
R2 of 0.61. The 1994 digitized photo generated NDVI values and pruning weights had an R2 of
0.91. We surmise this indicated that scanned aerial photography may produce similar results to
digitally acquired imagery if used with the classification procedure outlined above. A difference
image was also computed for these photographs (Figure 7) with patterns similar to those in
Figure 6, but with a higher spatial resolution.
Proximity Analysis
Areas that were conspicuously damaged by phylloxera for each year were identified on
the imagery to determine an NDVI class threshold and generate a vegetation stress image as
follows. For each year the classified NDVI image was combined with the corresponding false
color infrared (CIR) image to visually determine a threshold class number for 1993 and 1994.
Pixels below this threshold were considered to represent stressed vegetation in the NDVI
classified images. For the four blocks (C, H, I, and J), the same threshold was applied to each
based on similarities in age, vine and row spacing, and trellis type. The threshold may require
modification from block to block due to differences in vine canopy caused by various trellis
types, vine spacings, or age differences. For example, a block with wider row spacing would have
lower NDVI values than a block with narrower rows, due to the increase in soil area being sensed,
although the vine canopy may exhibit similar health characteristics. To compensate for this
NASA Technical Memorandum 112218
15
difference, the threshold value may need to be lowered for a less densely planted block. The
classified images for each year were then each recoded into a binary image exhibiting only stressed
and non-stressed areas (parts a and b of Figures 8-11).
A proximity search, a GIS function, was performed for each of the 1993 stress images out
to 40 m from the edge of the stressed areas. In the resulting proximity image the value at each
pixel is the distance from the stressed areas, where areas inside the stressed areas have a distance
value of zero. Starting with a distance value of zero, the search image was iteratively recoded into
a series of binary images, where the pixels within the specified distance from the 1993 stressed
pixels were predicted to be stressed in 1994. This resulted in a phylloxera stress prediction for
the following year. The vineyard manager can use this predictive tool to prepare for lower yields
or plot eradication for that block.
Finally, the stressed areas at each distance were compared to each of the 1994 stress
images to determine the best predictive match. The recoded image was compared with the 1994
stress image by calculating the percentage of mismatched pixels, or error, at each distance. There
were two error components: commission and omission. This gave four occurrence possibilities,
since a given pixel could be stressed or non-stressed in each year. The two error components
were: (1) the area that was not predicted but was stressed (error of omission) and (2) the area
that was predicted but was not stressed (error of commission). These two counts were totaled
and divided by the total number of pixels in each block to derive the error estimates of phylloxera
spread prediction.
The distance of minimum error was the spread rate from 1993 to 1994 for that block. The
percentage errors for the four blocks are shown in Table 3. These data demonstrate that the
spread rate of the stressed areas was higher for the blocks that were already moderately stressed,
consistent with previous research (Wildman, 1983). The predicted stress images are shown in
NASA Technical Memorandum 112218
16
part c of Figures 8-11.
Table 3. Proximity analysis prediction error results for two moderately stressed blocks (C and
H) and two lightly stressed blocks (I and J)
Minimum total error
Block Percent Distance (m)
C 15.6 20
H 26.5 16
I 21.7 6
J 18.3 6
Red Edge Inflection Point
In addition to the NDVI analysis, the red edge inflection point (REIP) model has the
potential to indicate year to year change in plant stress. The REIP is the point of maximum slope
in the spectrum of a leaf, between the red absorbance well and the NIR reflectance plateau (Figure
1). The soil reflectance curve is generally monotonic. For a healthy leaf, the red spectral
absorption feature (well) is broader and the REIP shifts to longer wavelengths (towards the NIR),
as compared to a stressed leaf. The REIP is expected to be insensitive to the amount of canopy
cover and is more sensitive to leaf reflectance changes than NDVI.
Before computing an REIP image, the 45° oblique CASI 1993 image was registered to the
nadir acquired CASI image, so the derived REIP values could be compared to plot related values.
The NDVI values computed from this registered image had an R2 = 0.89 with the NDVI values
from the nadir image. This indicated the change in sensor configuration and look angle had little
effect on the NDVI values. Due to the different look angle some variation, caused by non-
NASA Technical Memorandum 112218
17
Lambertian reflectance from the canopy was expected. REIP images were generated by fitting
pixel values from channels 4 - 8 (680 - 788 nm) to a third order polynomial and solving for the
location of maximum slope (Baret et al., 1992). The following form for radiance as a function of
wavelength was used for curve fitting:
L(λx) = a λx3 + b λx2+ c λx + d,
where L(λx) is the radiance in channel x centered at wavelength λx. The set of equations for all of
the channels was then solved for the coefficients a, b, c, and d at each pixel. Taking the second
derivative and setting it to zero resulted in the following equation for the REIP wavelength, REIP:
REIP = -b / 3a
The REIP wavelengths were calculated using radiance values for each pixel. For example,
given radiance values 33, 50, 78, 87, and 88 W/m2.sr.µm, for the channels centered at 681, 710,
737, 747, and 788 nm respectively, the equation for radiance was estimated using least squares
regression:
L(λx) = -0.00017557 λx3 + 0.38163 λx2 - 275.45 λx + 66077,
where the REIP was 724.55 nm.
The mean REIP's for the plots are shown in Table 4 and the REIP image is shown in
Figure 12. Though the REIP images were noisy, the coefficients of determination with mean
REIP per plot were high: 0.72 with pruning weight, 0.83 with oblique NDVI, 0.89 with nadir
NDVI, and 0.88 with (nadir) NDVI class. The stronger relationship between pruning weights and
REIP, versus pruning weights and NDVI, may be due to the difference in viewing geometry or
the increased sensitivity of the REIP.
NASA Technical Memorandum 112218
18
Table 4. Mean Red Edge Inflection Point (REIP) wavelengths from 1993 oblique CASI estimated
from five (680 - 788 nm) channels
Plot REIP Standard
Deviation
1 723.29 2.2
2 722.25 2.6
3 723.47 2.0
4 724.97 2.4
5 725.64 1.2
6 724.33 1.9
7 724.92 1.4
8 724.67 2.2
9 725.19 1.1
R2 with pruning weight 0.72
R2 with NDVI class 0.88
The narrow range of REIP values (3.4 nm) and the high standard deviations (mean 1.9 nm
for the nine plots, Table 4) led to an investigation of the relationship between noise in the
radiance values and the REIP. A one, two, and five percent error in radiance for each of the five
CASI channels was simulated and the REIP was calculated for each case. This analysis showed a
0.5 nm REIP shift per percentage error. The radiometric accuracy of the CASI sensor is about
two percent (Babey and Soffer, 1992), or a one nanometer error in the REIP due to instrument
calibration uncertainty. This helps explain the noisy appearance of Figure 12.
Thermal Imagery
NASA Technical Memorandum 112218
19
Due to thermal calibration problems with the AIRDAS sensor in the ambient temperature
channel (channel 4), brightness temperatures could not be extracted. An image of the block I
digital numbers is shown in Figure 13. Areas indicating stressed vegetation in this image were
similar to those in the 1994 NDVI classified image (Figure 3), but the patterns were not as clear.
Field measurements showed a small (< 1°C), although statistically significant, temperature
difference between infested and uninfested grapevine leaves, but a large difference between soil
(55-60°C) and grapevine leaves (26-30°C). The AIRDAS data had a spatial resolution of 5m, so
pixels represented a mixture of grapevine canopy and soil. Therefore, differences in temperature
represented in the image mostly corresponded to varying mixtures of soil and canopy rather than
plant temperature differences.
Conclusions
To determine phylloxera infestation and predict future vine stress using remotely sensed
data, a cooperative technology testing and development project involving a number of
participants was initiated. During this project, various digital imaging systems were flown over
test sites in Napa, California to determine the most efficient method and data available to locate
and predict phylloxera vine stress. In the course of the GRAPES project a number of different
sensors were flown. Because grapevines are a row crop, pixel sizes smaller than the row spacing
can obscure the spectral differences and make block patterns difficult to interpret. Considering
the image analysis performed, entire vineyards can be processed and patterns within blocks
observed with a spatial resolution of about three to five meters. This project acquired imagery
collected over an entire vineyard, but focused on one block or several blocks to verify the results.
The smallest vineyard-unit managers consider for many decisions is the block (about 1-10 ha), so
within block variability is superfluous. Because the cost of acquisition and processing remotely
NASA Technical Memorandum 112218
20
sensed digital imagery is inversely proportional to the pixel size, three to five meters is a good
tradeoff between cost and spatial resolution.
The use of the same sensor for each flight, while not necessary, would greatly simplify
data processing and improve confidence in the derived image products. If the spatial resolution
between sensors is different, the resolution of the sensors has to be matched by image analysis
techniques. Spectral values between different systems do not have to match; as long as the data
sets are not significantly different, because clustering each NDVI image and the NDVI itself
compensates for differences in atmospheric conditions.
The mean spectral-class value per plot was found to be highly correlated to the pruning
weight per plot for each year. The correlation between pruning weight and mean NDVI class per
plot was higher than the correlation between pruning weight and mean NDVI value per plot (R2 =
0.76 compared to 0.66). Good results with digitally acquired imagery were achieved without
sensor calibration and atmospheric correction by using spectrally classified data to examine
relative differences in canopy cover per year. Preliminary results indicate this classification
procedure should provide improved results with scanned photography as well as digitally
acquired imagery.
Histograms of the spectrally classified, digitally acquired, images for block I also showed
a change from a homogeneous block in 1993, with a sharply peaked histogram, to a relatively flat
histogram, non-uniform block, in 1994. If an unsupervised classification procedure is used, then
the classified images can be ground truthed or compared to other imagery to determine the
relationship between classes, canopy cover, and damage level. This information can be used to
determine a threshold for classes of stressed or conspicuously damaged vegetation. The effect of
different trellising types or row spacing on NDVI was not explicitly investigated, but the
approach should provide a good indication of relative differences within a given field.
NASA Technical Memorandum 112218
21
Proximity analysis provided a method of estimating the next year's phylloxera damage.
Given some initial conspicuously damaged areas within a block, a proximity spread, based on
region growing from the existing clusters, can provide an estimate of future damage. Since new
phylloxera infestation locations within a block occur, in addition to spreading from an existing
infestation location, and the spread rate within a block was not the same across a block,
prediction error was approximately 20%. For block I (a five-hectare block), the damaged area was
1.1 ha in 1993, 2.2 ha in 1994, and the estimated damaged area in 1994 was 1.8 ha, with an error
of about one-hectare. Phylloxera are also usually well established by the time a damaged area is
large enough to be considered a center of growth. By the following year the block will be
moderately or heavily damaged, and combined with the large prediction error, a prediction
beyond one year is not practical. Spread analysis does, however, provide a tool for exploring
different scenarios of spread rates and growth centers. Insufficient testing was done to determine
if a single growth rate could be uniformly applied to vineyard blocks and still obtain reasonably
accurate predictions, but the results of this study suggest different growth rates are needed.
Red edge inflection point (REIP) results and thermal imagery for the plots were
promising, but less meaningful than the NDVI products. The REIP results agreed well with the
pruning weights and NDVI class results, but they require special narrow spectral bandpass filters
along the red edge that are not commonly available. Due to the narrow range (a few nanometers),
REIP analysis also requires detailed spectral and radiometric calibration throughout the image.
This is a current problem with CCD (charge coupled device) arrays, and the data need to be
stable if multi-temporal studies are to be practical. Patterns, indicating stressed vegetation, were
present in the thermal imagery, but were not as distinct as in the NDVI imagery. Leaf
temperature was difficult to measure due to the open canopy. As an indication of canopy cover
and vegetative health, the easily derived NDVI and following classified imagery were easier to
NASA Technical Memorandum 112218
22
interpret than the REIP or the thermal imagery.
Commercial airborne imagery acquisition services exist to provide the data needed for
generating NDVI products. Only a red and a near infrared channel are needed to compute an
NDVI, but multiple channels in the red to near-infrared spectral region are needed for computing
the REIP. A thermal sensor is needed to acquire thermal imagery. In the next few years (1998)
satellites will provide commercial multispectral data with four-meter spatial resolution. Through
procedures such as those outlined here, airborne imagery acquired in the visible and near-infrared
can complement vineyard managers' knowledge gained from conventional ground-based
techniques, aerial photography, and experience.
The imagery indicated plant stress due to phylloxera and other sources, such as water
stress. Airborne imagery can serve multiple roles in vineyard management. Multi-year imagery
can be used to help identify the type of stress if the growth pattern can be identified. The
benefits versus the costs of multiple flights per year are still unknown, but the information gained
from a single flight per year was considered worthwhile to the Mondavi vineyard management
team. Knowledge of the pattern change from year to year allows the vineyard manager to
intervene and apply remedial measures as well as provide data for financial forecasts.
Recommendations
To monitor the phylloxera infestation of a vineyard, digital multispectral imagery, should
be acquired at least once a year at full canopy, between mid-season and harvest. The imagery
should have a spatial resolution not to exceed three meters and the use of the same sensor
package for each data collection period is important. The data can be used to generate co-
registered, classified, NDVI data sets for multi-year comparisons. In classifying one block, five or
six spectral classes are sufficient, but ten to twelve should be used for an image of an entire
NASA Technical Memorandum 112218
23
vineyard. A small number of classes makes image product interpretation easier, but some features
could be missed with too few classes. To avoid problems with changing amount of bare soil in the
imagery, a threshold should be used to eliminate low NDVI pixels.
Computer Facilities
All image processing described above was performed at the NASA Ames Research
Center's ECOSAT Computational Facility with ERDAS Imagine 8.20 (ERDAS, Inc., Atlanta,
GA) running under Solaris on a Sun Microsystems SPARCstation. The processing described here
used standard image processing routines available on a PC using any one of number of commonly
available geographic image processing software packages. More information about satellite image
processing, much of which also applies to aircraft imagery, including software sources, can be
found in the Satellite Imagery Frequently Asked Questions (FAQ) at
http://www.geog.nottingham.ac.uk/remote/satfaq.html.
Acknowledgments
Work to establish the plot size and locations and phylloxera levels within the plots was
performed by E. Weber (University of California Cooperative Extension Napa County), J. De
Benedictis (UC Davis), R. Baldy, and M. Baldy (both California State University Chico). GPS
data were collected by C. Bell (Johnson Controls World Services, NASA Ames Research Center,
currently California Department of Forestry through UC Davis) and B. Osborn (UC Davis,
currently Glen Ellen Carneros Winery). Pruning weights were courtesy of Robert Mondavi
Winery (D. Bosch). A. Bledsoe, D. Bosch (both Robert Mondavi Winery), and P. Freese (Wine
Grow) provided guidance throughout the project. Other contributors included D. Peterson, J.
Salute, and V. Vanderbilt (all NASA Ames Research Center). The work described in this paper
NASA Technical Memorandum 112218
24
was performed at the NASA Ames Research Center under UPN 233-01-04-05 in fiscal years
1993 - 1995.
Project URL (Web Page Address)
http://geo.arc.nasa.gov/sge/grapes/grapes.html
NASA Technical Memorandum 112218
25
Appendix: Sensor Specifications
The sensors flown in the course of this project were summarized in Table 1, but this
appendix describes these sensors in more detail. The sensors were flown on aircraft platforms at
low, medium, and high altitudes and had a various spectral characteristics. Half of these were
charge-coupled device (CCD) sensors and the other half were scanners. Scanning sensors use a
linear array of photo detectors that measures the intensity of radiance within some wavelength
region as the sensor passes (or sweeps) over the landscape, while a CCD sensor uses an array of
detectors and takes a "snapshot" of the landscape.
Sensors measure radiant intensity at some wavelength range and have two types of
resolution: spatial and radiometric. The wavelength region of each sensor channel is determined
by the spectral response function of the wavelength filter, and the bandwidth of the filter at half
maximum value is the full width half maximum (FWHM). Spatial resolution was defined on page
5 as related to aircraft altitude, but this was just another way of describing a detector's
instantaneous field of view (IFOV), or angular width. Because a detector's IFOV is fixed, a change
in altitude changes the amount of landscape subtended by the detector. This quantity is
expressed in radians, where 1.0 r = π/180°, or usually in milliradians. The radiometric resolution
of these sensors was eight bits, except the CASI sensor with twelve bits and the AIRDAS sensor
with sixteen bits. Some of these sensors, like CASI, can be reconfigured for multiple purposes,
but Table A1 describes the configurations used with the GRAPES project.
NASA Technical Memorandum 112218
26
Table A1. Specifications airborne sensors used in the GRAPES project.
Sensor Channel centers
(nm) FWHM (nm) Sensor type Image size
(pixels) IFOV (mr)
ADAR 440 80 CCD 1500x1000 0.44
515 110
650 80
890 220
AIRDAS 645 70 scanner 720 2.62
1635 130
4550 300
9250 7500
CASI (4 ch) 550, 630, 680, 767 12 scanner 512 1.2
CASI (8 ch) 500, 550, 630, 680,
710, 737, 747, 788 12 scanner 512 1.2
DMSV 450, 550, 650, 750 25 CCD 740x578 0.72
EO Camera 680 10 CCD 1025x1280 0.25
720 5
735 5
750 11
775 10
NS001 TMS 489 61 scanner 700 2.5
566 74
665 64
839 143
1240 220
1640 140
2240 280
NASA Technical Memorandum 112218
27
11300 2000
RDACS 548, 650, 821 10 CCD 739x484 1.0
NASA Technical Memorandum 112218
28
References
Ambrosia, V. G., J. A. Brass, J. B. Allen, E. A. Hildum, and R. G. Higgins. 1994. AIRDAS,
development of a unique four channel scanner for natural disaster assessment. First
International Airborne Remote Sensing Conference and Exhibition, Strasbourg, France, 11-
15 September 1994.
Babey, S. and R. Soffer. 1992. Radiometric calibration of the compact airborne spectrographic
imager (CASI). Canadian Journal of Remote Sensing 18(4):233-242.
Baret, F., S. Jacquemoud, G. Guyot, and C. Leprieur. 1992. Modeled analysis of the biophysical
nature of spectral shifts and comparison with information content of broad bands.
Remote sensing of Environment 41:133-142.
Borstad Associates, Ltd. 1991. Low cost digital remote sensing using the Compact Airborne
Spectrographic Imagery. Sidney, British Columbia, Canada: Borstad Associates, Ltd.
Duda R. and P. Hart. 1973. Pattern Classification and Scene Analysis. New York:John Wiley and
Sons, Inc.
Granett, J., A. Goheen, and L. Lider. 1987. Grape phylloxera in California. California Agriculture
41(1):10-12.
Granett, J., J. De Benedictis, J. Wolpert, E. Weber, and A. Goheen. 1991. Deadly insect pest
poses increased risk to north coast vineyards. California Agriculture 45(2):30-32.
Johnson, L. F., B. Lobitz, R. Armstrong, R. Baldy, E. Weber, J. De Benedictis, and D. Bosch.
1996. Airborne imaging aids vineyard canopy evaluation. California Agriculture 50(4):14-
18.
Lyon, R. J. P. 1994. SpecTerra digital multi-spectral video image data formats. Stanford, CA:
SpecTerra Systems, Pty, Ltd.
Tucker, C. J. 1979. Red and photographic infrared linear combinations for monitoring vegetation.
Remote Sensing of Environment 8:127-150.
Wildman, W. E., R. T. Nagaoka, and L. A. Lider. 1983. Monitoring spread of grape phylloxera by
color infrared aerial photography and ground investigation. American Journal of Enology
and Viticulture 34(2):83-94.
NASA Technical Memorandum 112218
29
Figure 1. Example of leaf and soil spectra. The REIP is the point of maximum slope in the
spectrum of a leaf. A healthy leaf has a broader spectral absorption in the red (680 nm) and REIP
occurs at a longer wavelength, as compared to a stressed leaf.
NASA Technical Memorandum 112218
30
1994 red channel
1993 red 5x5 1993 red 7x7 1993 red 9x9
Figure 2. Low pass filter kernels of 5x5, 7x7, and 9x9 pixels were applied to the CASI (1993) red
channel data to resolution match the image to the EO Camera (1994) red channel image. After
visual comparisons, the 7x7 average 1993 image determined to be the best match. Features, such
as the two bright areas in the 1993 images, were too distinct in the 5x5 average image, and were
too blurred in the 9x9 average image.
NASA Technical Memorandum 112218
31
1993 1994
Figure 3. Block I classified NDVI 1993 and 1994 images, where the lowest NDVI values are
shown in black and the highest values in white, also shown in white are the boundaries of the
block and plots within the block. A large patch of low canopy cover vines can be seen in the
lower right, contrasting with the trees at the extreme right edge of the 1993 image. Multiple
patches of low canopy cover vines can be seen in the 1994 image. "Rainbow" colored images
were used during image analysis and initial products were colored by vigor level: brown for little
or no vegetation, yellow for some vegetation, and green for high vegetation cover.
NASA Technical Memorandum 112218
32
Figure 4. NDVI class histograms for block I. Class 0 percentages represent pixels below the bare
soil threshold. Only ten of the twelve classes used for the ranch were present in block I.
Figure 5. Block I pruning weight per plot for both 1993 and 1994 correlated well with the per-
plot mean Normalized Difference Vegetation Index class number. In 1993, R2 = 0.60; in 1994, R2
= 0.88; and for both combined R2 = 0.76.
NASA Technical Memorandum 112218
33
Figure 6. Normalized Difference Vegetation Index class difference image for 1993 and 1994. This
image represents canopy cover change. Areas with the largest class value decrease are white and
those areas with a class value increase are black. The medium-gray area in the lower right
exhibited no change, also shown in white are the boundaries of the block and the plots within the
block.
Figure 7. Difference image between 1993 and 1994 NDVI classes derived from scanned
photography. Gray scale coding is the same as in the previous difference image. These one meter
resolution data were not precisely registered, so there was some misalignment of the vine rows.
The patterns visible in Figure 6 are still visible, however. If the images were smoothed and
resampled to 3m pixels, speckling can be reduced.
NASA Technical Memorandum 112218
34
a b c
Figure 8. Proximity analysis images for block C, part a) 1993 stressed areas, b) 1994 stressed
areas, and c) the 1994 stressed areas predicted at twenty meters using proximity to 1993 stressed
areas. White areas are stressed and black areas are unstressed or background.
a b c
Figure 9. Proximity analysis images for block H, part a shows 1993 stressed areas, b shows
1994 stressed areas, and c shows the 1994 stressed areas at sixteen meters predicted using
proximity to 1993 stressed areas.
NASA Technical Memorandum 112218
35
a b c
Figure 10. Proximity analysis images for block I, part a shows 1993 stressed areas, b shows
1994 stressed areas, and c shows the 1994 stressed areas at six meters predicted using proximity
to 1993 stressed areas.
a b c
Figure 11. Proximity analysis images for block J, part a shows 1993 stressed areas, b shows
1994 stressed areas, and c shows the 1994 stressed areas at six meters predicted using proximity
to 1993 stressed areas.
NASA Technical Memorandum 112218
36
Figure 12. REIP image of block I. The REIP was computed using five channels as input to a third
order polynomial. Some distortion of the image is apparent due to aircraft roll, but the boundary
of the block is visible and the dark areas (shorter REIP) correspond to stressed areas from the
1993 classified NDVI image (Figure 3).
Figure 13. Block I AIRDAS thermal image (Oct 1994), where the lowest digital numbers
(corresponding to the lowest brightness temperatures) are black and the highest white, also
shown in white are the boundaries of the block and the plots within the block. Patterns similar to
the 1994 NDVI classified image (Figure 3), but with low contrast within the block, and the gray
scale reversed.