Content uploaded by Eyal Ben-Dor
Author content
All content in this area was uploaded by Eyal Ben-Dor on May 26, 2017
Content may be subject to copyright.
© 2012 Science From Israel / LPPLtd., Jerusalem
Israel Journal of Plant Sciences Vol. 60 2012 pp. 77–83
DOI: 10.1560/IJPS.60.1-2.77
*Author to whom correspondence should be addressed.
E-mail: ilan@sensingis.com
Cannabis
Ilan azarIa,a naftalI GoldschleGer,b and eyal Ben-dora
aDepartment of Geography and Human Environment, P.O. Box 39040, Tel Aviv University, Tel Aviv 69989, Israel
bSoil Erosion Research Station, Ministry of Agriculture, Emek-Hefer 40250, Israel
(Received 10 October 2011; accepted in revised form 7 January 2012)
Honoring Anatoly Gitelson on the occasion of his 70th birthday
Drug use, mainly of Cannabis, has dramatically increased over the past two decades,
calling for efcient drug monitoring and prevention tools. This study evaluates the
use of ground-based hyperspectral detection for surveying and mapping Cannabis
cultivations. The ability to identify Cannabis plants at high spectral resolution us-
ing a ground-based hyperspectral detector (imaging spectroscopy sensor) outdoors
was measured using the AISA Eagle hyperspectral detector at 400–1000 nm wave-
lengths, at a distance of 75 m. Analysis of the measured data by image-processing
and statistical variation revealed that the spectral characteristics of Cannabis are
unique only within a wavelength range of 500–750 nm. It is important to notice that
variation was tested only with two species, and background was unique. Error in
classication (false alarm) was found between Cannabis canopy and citrus canopy:
15% of Citrus was classied as Cannabis.
Keywords: hyperspectral remote sensing, eld spectrometer, spectral resolution,
spatial resolution, Cannabis
Drug use, mainly of Cannabis, has dramatically in-
creased over the past two decades. Cannabis remains
the most widely used illicit substance in the world.
Globally, the number of people who used Cannabis at
least once in 2008 was estimated at between 129 and
191 million, or between 2.9% and 4.3% of the world
population aged 15–64 (World Drug Report 2007).
Drug prevention calls for accurate and updated informa-
tion on Cannabis elds, and the demand for monitoring
and detection tools that can cover large areas of drug-
oriented plants (e.g., Cannabis) has increased accord-
ingly. In this research, a ground hyperspectral camera
was used to assess the appropriateness of hyperspectral
technology for this task.
The main aim of our research was to develop hy-
perspectral methodology that will make it possible
to discriminate Cannabis crops from the surrounding
vegetation. This methodology was based on theoretical
knowledge acquired from studies analyzing the spectral
properties of plants, and on comprehensive measure-
ment of Cannabis canopy reectance. The hypothesis
was that the spectral properties of Cannabis leaves dif-
fer from those of other vegetation’s leaves according to
the chemical and geometrical properties of this plant.
Kalacska et al. (2007a) made use of hyperspectral
remote sensing technologies to map lianas growing
in the tropical forests. That study involved measure-
ments using an Analytical Spectral Devices (ASD) eld
spectrometer and a hyperspectral image with a spatial
Israel Journal of Plant Sciences 60 2012
78
resolution of 1 m and 210 spectral channels between
400 and 2500 nm. Kalacska et al. (2007b) used those
hyperspectral images to map the ecological variety of
plant life in tropical forests and to identify “ecologi-
cal ngerprints”, by employing models that determine
relationships between ground biomass data and various
hyperspectral indices. That study indicated the necessity
of collecting samples in different seasons (spectro-tem-
poral domain) to obtain a reliable ecological ngerprint.
In addition, they found that the best separation method
is wavelet decomposition.
A study conducted in California by Underwood et al.
(2003) showed that invasive habitats can be separated
from endemic plant life by hyperspectral means. Simi-
lar ndings were obtained by Asner et al. (2008), who
studied plant habitats in the Hawaiian Islands. Mapping
was performed with the 224-channel airborne visible
infrared imaging spectrometer (AVIRIS). Plant-life
characterization was carried out in a number of phases,
including atmospheric correction and image-processing
of the data to obtain reectance values. Specically, they
used an atmospheric-correction model, minimum noise
fraction (MNF) sorting to reduce imaging noise, princi-
pal component analysis (PCA) statistical processing to
separate the various plant groups, and spectral assort-
ment using continuum removal. Selected wavelengths
were in spectral ranges that allowed water-absorption
identication. MNF assortment was determined to be
the most successful method for isolating the various
groups.
Vrindts et al. (2002) showed the possibility of sepa-
rating seven families of weeds from corn and beet crop
plantations. Their images were obtained with a SPECIM
hyperspectral ground-based camera over the spectral
range of 400–1000 nm. That study was carried out un-
der laboratory conditions at a distance of 70 cm from
the object. The most signicant statistical sorting was
based on eight wavelengths situated in specic spectral
regions associated with the biochemical components
of the leaf. Despite the good results (94% separation
capacity between corn and weeds), it was concluded
that articial and imbalanced lighting decreases sorting
capacity.
Cannabis
Multispectral remote sensing
A study conducted by the United Nations’s Ofce on
Drug and Crime (UNODC) in Morocco used remote-
sensing methodologies to map Cannabis elds. In that
project, data were extracted only from multispectral sat-
ellite imagery and based on ground validation. The pro-
cess is complex, and requires a great deal of manpower
(Morocco Cannabis Survey, United Nations, 2003).
Spectroscopy and hyperspectral remote sensing
In California, spectral discrimination of the Can-
nabis plant from other plants using eld spectrometry
was studied (Daughtry and Walthall, 1998). In the rst
phase, spectral measurements were taken from Can-
nabis leaves and from other plants co-existing in the
same habitat. Measurements showed a signicant dis-
parity between the various leaves at the 550 nm (green
spectrum) and 720 nm (IR spectrum) wavelengths. Of
the wide range of plants measured, the lowest variance
was found between weeds and the Cannabis plant. The
second phase examined specic soil-related inuences
on the reectance spectrum of Cannabis leaf-tops in the
same region. Walthall and Daughtry (2001) examined
spectral reectance signatures of Cannabis leaves and
canopies using laboratory, eld, and airborne (AISA)
systems. They found that the spectral signature of Can-
nabis contrasts with other landscape signatures. The
spectral bands that included the most spectral separation
for Cannabis from other cover types were at 780, 800,
850, 880, and 900 nm using a Mahalanobis classier.
Cannabis crops were grown in the botanical garden of
Tel Aviv University under optimal irrigation and hu-
midity conditions (with special permission from the Is-
raeli Police Department and the Department of Health).
Plants were grown in pots on a peat–tuff substrate.
Cultivation was conducted in a naturally lit greenhouse,
and no fertilizers were used. Seed germination began
in mid-May, and two months later the rst owers ap-
peared. Hyperspectral measurements were conducted at
this phenological stage.
Spectral variance was gauged using a ground-based hy-
perspectral camera. Spatial variance between Cannabis
canopy, citrus, weeds, and green grass was examined
in unique resolutions, and the canopies of three species
dominate the measurement. The spectral separation
capacity was measured under full eld conditions using
only the AISA EAGLE camera. This phase was per-
formed in two corresponding experiments: Acquisition
from a 30-m high building, at an aerial distance of 60
to 75 m taken at 1130 h on 4 Apr 2008. Incidence angle
68º and sun angle was 67º. Pixel size was calculated to
be less than the size of one leaf, and the canopy of each
species was sampled randomly 25 times.
Azaria et al. / Identifying Cannabis using hyperspectral technology
79
The digital number (DN) recorded by the imaging sen-
sor for each spectral band was converted to radiance
and then to reectance units according to the supervised
vicarious calibration (SVC) method. This method relies
on calibrating imaging values in accordance with target
reection values (black and white bodies objects), gath-
ered in the imaging area using eld spectrometry. The
targets chosen for calibration are black nets of varying
density: 25%, 50%, 75%, and white (Brook and Ben-
Dor, 2011). This method decreases errors in the reec-
tance properties of each canopy species. High spectral
resolution data obtained from the AISA EAGLE sensor
are characterized by sensor noise and external factors
such as diffuse reectance that modify the spectral
properties of the leaves and the canopy’s reectance.
Moving average and Savitzky-Golay (Vaiphasa, 2006)
algorithms were used to smooth the hyper-spectral cube
obtained from ground acquisition.
For each one of the species a spectral library composed
of 25 pure pixels (spectral sampling) was built. Spectral
data were reduced and smoothed to 6 nm resolution.
Using that data, three mathematical methods were used
to identify spectral variance between vegetation species
(canopy scale) represented by these spectral libraries.
PCA was used to reduce the spectral data to small ma-
trices containing the most relevant information required
for discrimination. The rst derivative was calculated
on the continuum-removed spectra in the red-edge spec-
tral region (680–740 nm); continuum removal normal-
izes the reectance spectra, allowing comparisons of
individual absorption features from a common baseline
(Mutanga and Skidmore, 2007). Standardized variation
and standard deviation from the mean of the Cannabis
spectral signature (leaf and canopy level) were used to
score similarity between spectral signals. These three
methods are complementary and allowed us to reduce
errors in the classication. The aim of this step was to
localize the narrow spectral region in which the varia-
tion is repeatable and stable, and determine spectral
regions where Cannabis is distinct from other plant
species.
Looking at the post-processed reectance spectral sig-
nature of Cannabis, citrus, and grass canopies (Fig. 1)
extracted from ground hyperspectral imagery (75 m),
it can be seen that variation between species is deter-
mined by three spectral regions: VIS (550–690 nm),
NIR (700–720 nm, red edge), and 800–900 nm. Factors
that contribute to these variations are: the biophysical
attributes of the leaves, the orientation of the leaves,
leaf structure, canopy structure, soil reectance, illu-
mination and viewing conditions, and water content in
the canopy (Asner, 1998). Walthall et al. (2001) found
the Cannabis canopy bands showing the most spectral
separation from other cover types to be 780, 800, 850,
880, and 900 nm. To more closely analyze the variation
between the Cannabis canopy and other cover types,
three methods were used: standardized variation from
the mean of the Cannabis canopy spectrum, rst deriva-
tive of continuum removal of canopy reectance, and
PCA.
Cannabis
Internal spectral variation of the Cannabis canopy
Figure 2 shows the tendency toward dispersion
around the mean and median spectral signature of the
Cannabis canopy between 450 and 950 nm. The spec-
tral region with the largest variance was 740–950 nm,
with a standard deviation of between 0.035 and 0.068
in this region (Fig. 3). The large variation between 740
and 950 nm is explained by some authors as resulting
from photon scattering at the air–cell interfaces within
the leaf spongy mesophyll (Asner, 1998; Kuo-Wei et
al. 2005). In contrast, small variation was found in the
VIS and red-edge regions (690–740 nm). Asner (1998)
explained that the relatively stable optical properties of
the leaves at VIS wavelengths are due to biochemical
Fig. 1. Mean of spectral signatures of Cannabis, citrus, and
green grass canopies obtained from the ground hyperspectral
AISA Eagle camera.
Israel Journal of Plant Sciences 60 2012
80
characteristics resulting from the presence of biologi-
cally active pigments.
Spectral variation between species
Figure 4 illustrates the dispersion around the stan-
dardized deviation of the mean canopy spectral signa-
ture. In the VIS region between 500 and 640 nm, both
green grass and citrus canopy Z-score (distance from
0) values are negative and far from the Cannabis mean
(Table 1). It is important to note that between 750 and
950 nm, the citrus canopy and Cannabis canopy are
similar (–1 < Z-score < 0). In contrast, between 750 and
950 nm, the green grass canopy is highly heterogeneous
relative to the Cannabis canopy (Z-score > 2).
In the PCA of hyperspectral data acquired from 75 m
(Fig. 5), PCA1 explains only 62% and PCA2 26% of
the variation, mostly due to the resolution of the target.
In this study, the size of the target was constant, so we
assume that environmental factors reduced the ability
to discriminate Cannabis from the other species. The
major regions contributing to the spectral variation be-
tween species were 520–580 and 705–720 nm (n = 50,
the entire series is represented).
Figure 6 shows the rst-derivative reectance on con-
tinuum-removed spectra of Cannabis, citrus and green
grass canopies. The rst-derivative peak for the Canna-
bis spectrum was located at 705 nm, at 715 nm for citrus
Fig. 2. Dispersion tendency around mean and median spectral
signature of Cannabis canopy.
Fig. 3. Standard deviation of Cannabis canopy spectral signa-
ture (n = 25).
Fig. 4. Dispersion around the standardized deviation of mean canopy spectral signature.
Azaria et al. / Identifying Cannabis using hyperspectral technology
81
Table 1
Spectral variation range for each spectral region expressed as
Z-score values
Z-score Wavelength (nm) Species
–1.5 < Z < –3.8 515–617 green grass
–1.5 < Z < –3.4 700–724 green grass
1.5 < Z < 5.3 740–950 green grass
–2 < Z < –4.6 512–588 citrus
–1.7 < Z < –2.3 695–720 citrus
Fig. 6. First-derivative reectance on continuum removed spectra of Cannabis, citrus and green grass canopies. Cannabis curve
shifts to short wavelengths.
Fig. 5. PCA components 1 and 2, spectral libraries of foliage from Cannabis and citrus, retrieved by hyperspectral imaging at
75 m.
leaves and at 720 nm for green grass. Figure 7 shows a
low standardized variation of the median signature from
the mean for Cannabis, green grass and citrus. The box
plot in Fig. 8 explains the variation in the location of the
rst-derivative peak more clearly. Variation in the loca-
tion of the maximum peak in the red edge is related to
nitrogen concentration (Mutanga and Skidmore, 2007).
Hyperspectral remote sensing gives continuous spec-
tral signatures of materials. For green vegetation, this
technology enables extracting information on several of
the plant’s chromophores. Chlorophyll absorbs strongly
in the blue (450 nm) and red (670 nm) regions, reects
strongly in the NIR region (700–1300 nm), and shows
strong water absorption at around 1400 and 1900 nm
Israel Journal of Plant Sciences 60 2012
82
Fig. 7. Median variation around mean spectrum for citrus,
Cannabis, and green grass canopies.
Fig. 8. Box plot of rst-derivative peak for Cannabis, citrus,
and green grass canopies. First derivative of citrus spec-
trum central tendency is exactly at 715 nm; n = 25 for each
species.
(Govender et al., 2007). Although these are common
spectral properties for all green vegetation, a small
variation exists in the biophysical and biochemical sta-
tuses of plants (Jacquemoud et al., 1996). Moreover, the
structure of the mesophyll layer, the external structure of
the leaves, and the plant’s architecture are major factors
determining spectral variation among species (Zhumar,
1999; Walthall and Daughtry, 2001). From the results
obtained in this study, it was possible to see variations
only in the VIS–NIR region (512–580 and 705–720 nm)
as compared to weeds and citrus.
The spectral variation in these specic narrow
wavebands can be explained by the physiological and
biochemical variations between the species examined
(Cannabis, citrus, and grass). In general, photosynthetic
activities are not similar between species, and this may
be the case here. In addition, ecological and genetic
factors can also contribute to the abovementioned varia-
tions (Martin et al., 2007). However, although a spe-
cic investigation of the physiological and biochemical
variations of Cannabis was beyond the scope of this
study, our results enabled discrimination between the
examined species.
Three mathematical tools were used to analyze the
variation between groups: rst derivative, rst deriva-
tive of the continuum removal, and PCA. These meth-
ods are commonly used to discriminate between groups
in a highly clustered environment. Although there are
other applicable methods, the success of the methods
used herein in discriminating between Cannabis, citrus,
and weeds suggests that the spectral information in the
VIS region is reliable and solid, and enables locating
errors and trends. For example, in Fig. 5, 15% of citrus
was classied as Cannabis, and Fig. 4 shows this simi-
larity as a distance (Z-score) from the mean of the Can-
nabis spectrum. The origin of this error is explained by
similarity between Cannabis and citrus in the 720–950
nm spectral region. This region is characterized by inter-
nal scattering of the mesophyll layer and is not favorable
for classication, even if it is statistically signicant. In
general, we can say that despite some restrictions in the
spectral discrimination of green vegetation across the
VIS–NIR–SWIR spectral region and the spectral simi-
larity of Cannabis to the other plants tested, signicant
information could be culled from the major chlorophyll
bands enabling spectral detection of Cannabis.
This research was funded by the Israel Anti-Drug Au-
thority.
Asner, G.P. 1998. Bioxphysical and biochemical sources of
variability in canopy reectance. Remote Sens. Environ.,
64: 234–253.
Asner, G.P., Jones, M.O., Martin, R.E., Knapp D.E., Hughes,
F.R. 2008. Remote sensing of native and invasive species
in Hawaiian forests. Remote Sens. Environ. 74: 69–84.
Brook, A., Ben-Dor, E. 2011. Supervised vicarious calibration
(SVC) of hyperspectral remotesensing data. Remote Sens.
Environ. 115: 1543–1555.
Daughtry, C.S.T., Walthall, C.L. 1998. Spectral discrimination
of Cannabis sativa L. leaves and canopies. Remote Sens.
Azaria et al. / Identifying Cannabis using hyperspectral technology
83
Environ. 64: pp. 192–201.
Govender, M., Chetty, K., Bulcock, H. 2007. A review of hy-
perspectral remote sensing and its application in vegetation
and water resource studies. Water SA 33: 145–152.
Jacquemoud, S., Ustin, S.L., Verdebout, J., Schmuck, G., An-
dreoli, G. and Hosgood, B. 1996. Estimating leaf biochem-
istry using the PROSPECT leaf optical properties model.
Remote Sens. Environ. 56: 194–202.
Kalacska, M., Bohman, S., Sanchez-Azofeifa, G.A., Castro-
Esau, K., Caelli, T. 2007a. Hyperspectral discrimination
of tropical dry forest lianas and trees: comparative data
reduction approaches at the leaf and canopy levels. Remote
Sens. Environ. 109: 406–415.
Kalacska, M., Sanchez-Azofeifa, G.A., Rivard, B., Caelli, T.,
White, H.P., Calvo-Alvarado, J.C. 2007b. Ecological n-
gerprinting of ecosystem succession: estimating secondary
tropical dry forest structure and diversity using imaging
spectroscopy. Remote Sens. Environ. 108: 82–96.
Kuo-Wei, C., Yuan, S., Jeng-Chung, L. 2005. Predicting rice
yield using canopy reectance measured at booting stage.
Agron. J. 97: 872–878.
Martin, R., Asner, G., Sack, L. 2007. Genetic variation in
leaf pigment, optical and photosynthetic function among
diverse phenotypes of Metrosideros polymorpha grown in
a common garden. Oecologia 151: 387–400.
Morocco Cannabis Survey, 2003. United Nations: www.
unodc.org/unodc/en/crop_monitoring.html
Mutanga, O., Skidmore, A.K. 2007. Red edge shift and
biochemical content in grass canopies. ISPRS Journal of
Photogrammetry and Remote Sensing 62: 4–42.
Underwood, E., Ustin, S., Dipietro, D. 2003. Mapping non-
native plants using hyperspectral imagery. Remote Sens.
Environ. 86: 150–161.
Vaiphasa, C. 2006. Consideration of smoothing techniques for
hyperspectral remote sensing. Journal of Photogrammetry
and Remote Sensing 60: 90–99.
Vrindts, E., De Baerdemaeker, J., Ramon, H. 2002. Weed
detection using canopy reection. Precision Agriculture
3: 63–80.
Walthall, C.L., Daughtry, C.S.T. 2001. Detecting marijuana
using optical wavelength remote sensing. In: Proceedings
of 2001 ONDCP International Technology Symposium,
25–28 June 2001, San Diego, CA, pp. 648–656.
Zhumar, Y. 1999. Effect of the structure of mesophyll in a
leaf on the form of the spectral curve for the rst deriva-
tives of the reection coefcients of leaves over the range
0.68–0.75 µm. J. Appl. Spectrosc. 65: 967–972.
Unodc: http://www.unodc.org/unodc/data-and-analysis/
WDR.html
World Drug Report, 2007. United Nations, NY.
