Content uploaded by Francisca López-Granados
Author content
All content in this area was uploaded by Francisca López-Granados on Jul 09, 2019
Content may be subject to copyright.
DOI: 10.1111/j.1365-3180.2010.00829.x
Weed detection for site-specific weed management:
mapping and real-time approaches
FLO
´PEZ-GRANADOS
Institute for Sustainable Agriculture ⁄CSIC, Co
´rdoba, Spain
Received 22 December 2009
Revised version accepted 6 September 2010
Summary
This work describes the current status of remote and
proximal (on-ground) weed detection systems for site-
specific weed management and discusses the limitations
and opportunities of these technologies. Remote sensing
based on multispectral aerial imagery can provide
accurate weed maps, especially at late weed phenological
stages, whereas images from high spatial resolution
satellite and unmanned aerial vehicles must still be
analysed. Hyperspectral images produce highly accurate
maps at early and late phenological stages at a farm
scale or medium spatial scale. However, this technology
is not profitable, because of current operating costs,
which are prohibitive. In studies of on-ground weed
seedling detection, accurate results can be obtained at a
medium farm scale. Despite numerous efforts, a power-
ful and flexible classifier of soil, weeds and crops in a
number of situations, remains the greatest challenge of
this technology. The main limitations of remote and
proximal sensing may be summarised in the following
two points: (i) the time and education required for
applying new technological advances and (ii) the high
cost of the technology and the lack of compatibility of
the machinery. Possible solutions might include: (i)
offering an advisory service that provides technical
support, agronomic knowledge and specific training
courses, (ii) the development and implementation of
uniform and cheaper standards, (iii) increased research
of both high resolution satellite imagery exploring
object-based image analysis and pan-sharpened imagery
and unmanned aerial vehicles (UAV) and (iv) enabling
the development of current prototypes of robotic
weeding into commercial products. The general lack of
multidisciplinary research groups can be a disadvantage
when comparing the economic feasibility of site-specific
weed management with conventional systems.
Keywords: aerial and satellite imagery, hyperspectral,
mapping weeds, multispectral, neural networks, preci-
sion agriculture, robotic real-time, remote sensing,
spectral signatures, machine vision.
L
O
´PEZ
-G
RANADOS
F (2011). Weed detection for site-specific weed management: mapping and real-time approaches.
Weed Research 51, 1–11.
Introduction
Site-specific weed management: general concepts
Site-specific weed management (SSWM) includes spray-
ing weed patches only and ⁄or adjusting herbicide
applications according to weed density or weed species
composition (e.g. herbicide resistant, broad-leaved or
grass weeds). This strategy fits well with European goals
of minimising herbicide use and tracing farm products
(FP7-NMP, 2009). In many parts of Europe, field sizes
are increasing, farmers frequently have a university
degree and usually manage several fields (sometimes
more than 3000 ha in southern Spain). These factors
favour the adoption of precision agriculture, because the
investments in technology and human skills can produce
a faster and more positive return (Reichardt & Ju
¨rgens,
2009). SSWM represents a four-step cyclical process
Correspondence:FLo
´pez-Granados, Institute for Sustainable Agriculture ⁄CSIC, P.O. Box 4084, 14080-Co
´rdoba, Spain.
Tel: (+34) 957 499 219; Fax: (+34) 957 499 252; E-mail: flgranados@ias.csic.es
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
that includes (i) weed monitoring consisting of ground
sampling or detection of weeds, (ii) decision-making
(also called management planning), referring to the
design of an action based on the diagnosis and other
available information (e.g. previous farmer experience),
(iii) precision field operation, which is the execution of
SSWM and (iv) evaluation of the economic profitability,
safety and environmental impacts of the field operations
for the next season. Weed monitoring in crops is still one
of the critical components for the adoption of SSWM.
Large research efforts have been focused on this step,
which is a prerequisite for both map-based and real-
time-based weed monitoring.
Weed monitoring: general concepts
This article aims to review recent advances in monitor-
ing of arable (in crop) weeds for SSWM (mainly for
herbicide control) by considering remote (airborne-,
satellite- and unmanned-based platforms) and proximal
(ground-based sensors and cameras) sensing. The focus
will also be on the opportunities and limitations of the
application of this technology and on the need for future
investigations into increase the acceptance and adoption
of SSWM.
Site-specific weed management with a map-based
approach consists of developing robust methods for
weed data acquisition, analysis and delineation of
management zones for further use. The spatial distribu-
tion of weeds within a crop can be detected and mapped
by using remote sensing, which includes aerial and
satellite imagery. It can also be detected by proximal
sensing (Srinivasan, 2006), which refers to in-field
machine-mounted (harvesters, tractors, robots) sensors.
In contrast, real-time monitoring and spraying consists
of a weed control system that can simultaneously detect
and control weeds. This strategy requires monitoring
processing techniques, decision-making and spraying
while the vehicle is moving forward at a constant speed
(e.g. Lee et al., 1999). Therefore, remote sensing is
helpful for map-based SSWM, whereas proximal sensing
is useful for both.
Remote sensing
The spectral response of plant species at the canopy or
single-leaf scale is unique and is known as spectral
signature. One of its defining characteristics is that it
varies according to the phenological stage and it can be
measured by proximal or remote sensors (Pen
˜a-Barraga
´n
et al., 2006; Lo
´pez-Granados et al., 2008). The basic
principle is that if differences in reflectivity based on
external factors or distinctive phenological stages can be
measured or recognised, there is potential for automatic
weed detection or mapping (monitoring). Multispectral
broadband sensors usually collect data for 3–7 bands
of around 100 nm width, whereas hyperspectral sensors
detect many narrow and contiguous wavelengths,
usually <10 nm width (Table 1). Thus, as bandwidths
are narrower in hyperspectral scanner systems, small
variations in reflectivity can be detected that might
otherwise be masked within the broader bands of multi-
spectral scanner systems.
Together with the importance of spectral resolution,
the other essential parameter in remote sensing is to
select the suitable pixel size, based on the inherent
properties of the input data, i.e. what is the smallest
discernable feature at any given spatial resolution and
the accuracy at which it is mapped. Hengl (2006)
discussed the rules of thumb to find the right pixel size
to help inexperienced users select the appropriate spatial
resolution. He concluded that at least four pixels are
required to detect smallest objects and at least two pixels
to represent the narrowest objects, being objects the
smallest size area that we map (weed plants or weed
patches in our case). In other words, if the smallest
object is a weed patch of 16 m
2
(4 ·4 m), we should use
imagery with resolution of 2 m and finer. In Table 1 are
shown some of the current multispectral and hyperspec-
tral sensors, with the appropriate spatial resolution to
detect and map weeds for SSWM. In the following
sections, the detection of seedling or mature weeds based
on the use of imagery with different spectral and spatial
resolutions is presented.
Early (seedling) weed detection
In most weed control situations including SSWM, it is
generally necessary to control weeds at an early growth
stage of the crop. However, remote sensing of grass
weed seedlings in monocotyledonous crops and seed-
lings of broad-leaved weeds in dicotyledonous crops
presents much greater difficulties than mapping them in
the late stage, for three main reasons: (i) cereal crops and
grass weeds (and also many dicotyledonous crops and
broad-leaved weeds) generally have similar reflectance
characteristics early in the season, which could involve
the necessity of using hyperspectral data to detect small
variations in reflectivity, (ii) the distribution of weeds
can be in small patches, which could indicate the
necessity to work with very high spatial resolution
imagery and (iii) the soil background reflectance inter-
feres with detection (Thorp & Tian, 2004).
The next point is to determine whether the criterion
for SSWM might be the presence or absence of weeds, or
differentiation into monocotyledonous or dicotyledon-
ous groups without any differentiation of weed species,
as these weed groups are treated with different herbicides
2FLo
´pez-Granados
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
and usually have different spectral characteristics. There
are only a few reports on the ability of airborne
multispectral imagery to create accurate weed seedling
maps in crops for in-season site-specific post-emergence
herbicides. Brown and Steckler (1995) discriminated
several broad-leaved and grass weeds in early-stage
no-tillage maize (Zea mays L.) using 10-cm pixels. Lamb
et al. (1999) mapped continuous Avena spp. populations
in a seedling triticale crop using 0.5- to 2.0-m pixels.
Medlin et al. (2000) discriminated seedlings of several
weed species in early-season soyabeans (Glycine max L.)
with 1-m spatial resolution. These works concluded that
further investigations should evaluate the accuracy of
weed detection with high spatial resolution imagery
(pixels of at least <0.5 m). Gray et al. (2008) discussed
the accuracy of weed seedling maps representing bare
soil, soyabean and six weed species grouped. They also
examined the accuracy for bare soil, soyabean and all
weed species independently, using aerial images with
50-cm pixels taken 8 and 10 weeks after emergence.
They reported that the accuracy was higher when remote
sensing data were obtained 8 weeks after emergence and
that the differentiation of each weed species would
require hyperspectral airborne imagery and a more
powerful image analysis technique than the maximum-
likelihood classifier. This imagery has the appropriate
spectral resolution to detect weeds at the early stage.
However, very little literature is available on the use of
hyperspectral airborne imagery for weed detection
during the early growth stages of a crop. Martı
´net al.
(2009) mapped three weed seedling species (Sorghum
halepense L., Xanthium strumarium L. and Abutilon
theophrasti Medicus) in early maize using an Airborne
Hyperspectral Scanner (AHS) with 80 wavelengths in
the visible and near-infrared domains. Goel et al.
(2003a,b) and Karimi et al. (2005) used CASI (Compact
Airborne Spectral Imager) hyperspectral sensors with 72
narrow bands (also visible to near-infrared) to map grass
weeds (Cyperus esculentus L., yellow nut sedge; Echino-
chloa crus-galli L. Beauv., barnyard grass; Digitaria
ischaemum Schreb., crabgrass) and broad-leaved weeds
(Cirsium arvense L. Scop., creeping or Canada thistle;
Sonchus oleraceus L., sow thistle; Amaranthus retroflexus
L., redroot pigweed; Chenopodium album, fat-hen,
among others) in maize fields. Despite this research,
the main restriction of hyperspectral remote sensing is
Table 1 Some of the current sensors and platforms with spatial and spectral resolutions required for site-specific weed management
Sensors and platforms
Spatial
resolution (m)
Waveband
interval (nm) Altitude (km) Revisit time (days)
Multispectral satellite
IKONOS 4* 450–900
681 1.5
QuickBird 2.44* 450–900
450 1–3.5
GeoEye-1 (former OrbView5) 1.64* 450–920
à
681 2.1–2.8
Airborne
Daedalus 1268 3.44
§
420–13000
–
1.37
Conventional turboprop 0.30 400–900** 1.52
Twin-engine plane
Unmanned aerial vehicle (UAV) 0.15
§
490; 530; 570; 670; 700; 750; 800
0.15
Hyperspectral airborne
AVNIR
#
1 430–1012 (10 nm) 1.5
CASI 1–3
§
400–1000 (1.9 nm) 0.84–3.5
AHS 2–3.44
§
430–12500 (13–300 nm) 1–1.37
HyMap 2 450–2500 (20–10 nm) 2
AVIRIS 4 400–2500 (10 nm) 3.8
*1, 0.61, and 0.5 m spatial resolution in Panchromatic for IKONOS, QuickBird and GeoEye respectively.
Bands: Blue, 450–520; Green, 520–600; Red, 630–690; Near-infrared, 760–900.
à
Bands: Blue, 450–510; Green, 510–580; Red, 650–690; Near-infrared, 780–920.
§
Spatial resolution depends on flight altitude and camera field of view (FOV). Some examples as follows: angular FOV of 85.92and
1.376 km flight altitude yield 3.44 m pixel for Daedalus and AHS; angular FOV of 42.8·34.7and 0.150 km flight altitude generate
0.15 m pixel size for UAV; angular FOV of 60and 2 km flight altitude yield 2 m pixel for Hymap. For Hyperspectral imagery, pixel size
can also depends on the program to capture the image: for example CASI can offer submetre (0.7 m) pixel when only several wavebands
(e.g. 18 wavelengths rather than the 288 available) are selected.
–
Fixed channels of Daedalus 1268: 420–450; 450–520; 520–600; 600–620; 630–690; 690–750; 760–900; 910–1050; 1550–1750; 2080–2350;
8500–13000.
**
Bands: Blue, 400–500; Green, 500–600; Red, 600–700; Near-Infrared, 700–900.
These channel centres are just an example because different bandsets can be selected depending on the objectives adopted for the UAV
study.
#
AVNIR, Airborne Visible and Near-Infrared; CASI, Compact Airborne Spectrographic Imager; AHS, Airborne Hyperspectral Scanner;
AVIRIS, Airborne Visible ⁄Infrared Imaging Spectrometer.
Weed detection for site-specific weed management 3
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
that it is only cost-effective on a large scale, or if two
objectives, such as site-specific herbicide and fertilisation
applications, are solved in the same operation.
Currently, the multispectral satellites with higher
spatial resolutions with potential for discriminating
weed seedling in crops are QuickBird and GeoEye, with
2.44- and 1.64-m pixels in multispectral resolution
respectively. QuickBird has provided accurate maps of
C. arvense in sugar beet at the cotyledon stage (Backes &
Jacobi, 2006), but no information has been found
regarding the use of GeoEye imagery for mapping
weeds in the early growth stages of crops.
Late (mature) weed detection
An alternative to early detection is late detection,
because the differences in the weed-crop life cycle at
advanced phenological stages may enhance spectral
differences. Thus, detection of late-season weed infesta-
tion has considerable possibilities when the soil surface
is completely covered, the weeds exceed the crop canopy,
and the spectral differences between crops and weeds are
present and are quantifiable. Because weed infestations
can be relatively stable in location from year to year
(Barroso et al., 2004; Heijting et al., 2007a; Jurado-
Expo
´sito et al., 2004, 2005), late-season weed detection
maps can be used to design SSWM in subsequent years,
to apply in-season post-emergence herbicides if adequate
pre-emergence control was not achieved, or to know
where the greatest weed seed rain and seedbanks are in
the field (Koger et al., 2003). Post-emergence site-
specific applications can be useful to control broad-
leaved and grass weeds in maize (Brown & Steckler,
1995), to control cruciferous (e.g. Sinapis spp. and
Diplotaxis spp., De Castro et al., 2009) or grass weeds
(e.g. Avena spp.) with specific and very expensive
herbicides in cereals, or to treat herbicide-resistant
weeds such as ryegrass (Lolium rigidum Gaudin)
(Lo
´pez-Granados et al., 2008). Thus, the discrimination
of these problematic weeds at advanced phenological
stages could justify the necessity in reducing herbicide
use and cost, or of planning a specific application for
herbicide-resistant weeds.
Conventional colour (400–700 nm) and colour-infra-
red (500–900 nm) films for aerial photography have
been reported to accurately map late infestations of
maize caraway (Ridolfia segetum Moris) in sunflower
with 0.40-m pixels (Pen
˜a-Barraga
´net al., 2007). Gibson
et al. (2004) used aerial multispectral digital images with
1.3-m pixels to map late infestations of weeds in
soyabean while considering both the effects of weed
density and weed species discrimination accuracy. They
reported that classification accuracy decreased as weed
density or cover decreased. This finding indicates that a
greater spatial resolution is required to detect low weed
densities and that weedy pixels (presence of weeds) were
separated from weed-free and bare ground (absence of
weeds) with maximum 11% error, whereas the classifi-
cation error of weed species ranged from 17% to 39%.
Lo
´pez-Granados et al. (2006), using greater spatial
resolution aerial imagery (0.30-m pixel), reported a high
accuracy (ranging from 85% and 90%) when no weed
species were taken into account. However, they did not
improve the weed species classification (accuracy
<75%) when discriminating between wild oat (Avena
sterilis L.), canary grass (Phalaris brachystachys Link.)
and L. rigidum in wheat. To overcome this limitation,
the use of hyperspectral imagery or a more powerful
image analysis algorithm than vegetation indices or
Spectral Angle Mapper methods have been suggested to
improve classification results. Gutie
´rrez-Pen
˜aet al.
(2008) applied neural networks and improved the
accuracy of previous R. segetum maps (Pen
˜a-Barraga
´n
et al., 2007) by up to 15%. No information has been
found regarding the use of hyperspectral imagery for
mapping late weeds in crops.
Recently, unmanned aerial vehicles (UAV), which are
unpiloted aircraft guided by autonomous navigation
systems, have been presented as a promising tool in
several agricultural applications (Schmale et al., 2008).
The development of UAV could have great potential to
monitor early and late weeds for site-specific applica-
tions, because they can provide low-cost near-real-time
approaches with high spatial, spectral and temporal
resolutions and low-cost autopilot systems. However, at
present, there is very little information about UAV and
weeds. Herwitz et al. (2004) mapped guinea grass
(Panicum maximum Jacq.) within coffee fields when the
weeds were yellow-green and the coffee trees were
darker-green. Go
¨ktog
˘an et al. (2010) have developed a
rotary-wind UAV to survey infestations of two aquatic
weeds (Alternanthera philoxeroides Mart.Griseb, alliga-
torweed; and Salvinia molesta Mitchell, salvinia, giant).
Information on UAVs and weeds is scarce, particularly
for arable cropping.
Commission vs. omission errors
The agronomic implications of underestimation and
overestimation of weed pressure is usually addressed
through studying the risk of omission and commission
errors. Omission error indicates the proportion of true
weedy pixels not identified in the detection method and
thus contributes to underestimation of the weed patches.
Commission error indicates the proportion of pixels
misclassified as weedy, meaning that the classification
overestimates the weed patches and therefore provides a
conservative estimation. From an agronomic point of
4FLo
´pez-Granados
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
view, it is more desirable to select a classification in
which the omission error tends to be lower and the
commission tends to be higher so that the weed patches
are less likely to be missed. Gibson et al. (2004), Lamb
and Weedon (1998) and Pen
˜a-Barraga
´net al. (2007)
obtained less omission than commission errors in their
weed detection studies in soyabean, oilseed rape stubble
and sunflower with multispectral aerial imagery.
Limitations and opportunities of remote
sensing
The most important findings of the previous sections
regarding early or late detection by remote sensing are
(i) the primary aim for detecting and mapping weeds at
early or late phenological stages is the presence or
absence of weeds and not the differentiation between
individual weed species, (ii) this is achievable with
acceptable accuracy with pixels around 0.5 m and visible
and near-infrared multispectral windows, and (iii) weed
discrimination decreases with poorer spatial resolution.
To discriminate weed species, it is necessary to explore
hyperspectral airborne data.
Moran et al. (1997), Lamb and Brown (2001) and
Shaw (2005) reviewed the potential for image-based
remote sensing to provide spatial and temporal infor-
mation for SSWM. They concluded that the potential
market for remote sensing products in precision man-
agement is good, focussing on the necessity of assimi-
lating remotely sensed information into efficient decision
support systems. The current challenge for SSWM with
multispectral remote sensing is to check whether high
spatial resolution images (pixel < 0.5 m) can accurately
map weed seedling in crops for in-season, operational
selective herbicide applications. This line of research can
be complemented or overcome by mapping weeds at
later phenological stages for the next season ⁄s or for
in-season post-emergence herbicide site-specific control.
The development of low-cost autopilot systems, along
with the availability of very small pixel size (e.g. 15 cm,
Table 1) and a reduction in the weight and price of
sensors that can be installed in the UAV, has meant that
scientific interest in this kind of platform is growing for
different site-specific agricultural applications. In the
case of weed detection, there is a wide and promising
area that requires research attention. The only price
constraint on imagery acquirement is the size of the
surveyed area.
Companies that distribute high spatial resolution
satellite imagery usually offer users two separate prod-
ucts: a high spatial resolution panchromatic image
(QuickBird, 0.7-m pixel; GeoEye, 0.5-m pixel) and a
lower spatial resolution multispectral image in the
visible and near-infrared spectral range (QuickBird,
2.44-m pixel; GeoEye, 1.64-m pixel), as well as average
revisits of 1–3.5 and 2.1–2.8 days for both QuickBird
and GeoEye imagery. The high spatial resolution of
panchromatic images would have the potential to
accurately map weed patches, and the high spectral
resolution of multispectral images would facilitate the
discrimination of weeds and crops. To obtain an image
of high spectral and spatial resolutions, a pan-sharpen-
ing process or image fusion can be used to provide a
pan-sharpened or single image that combines high
spectral information (from multispectral imagery) and
high spatial information from the panchromatic image.
As a result, a 4-band multispectral pan-sharpened image
with a spatial resolution of 0.7 or 0.5 m for QuickBird or
GeoEye can be obtained. Castillejo-Gonza
´lez et al.
(2009) used a wavelet-based fusion technique in Quick-
Bird imagery to improve the performance of several
classification methods for precise monitoring of crops.
The use of this fusion technique could be a potential
solution to improve the accuracy of maps of early or late
weeds in crops.
Another way to improve the accuracy of weed maps
is to group adjacent pixels into spectrally and spatially
homogeneous objects created via a segmentation pro-
cess. Blaschke (2010) gives an overview of the develop-
ment of object-based methods and suggests that the
pixel paradigm is beginning to show cracks because of
object-based methods represent a significant new trend
in remote sensing for many monitoring programmes.
The objects are not characterised by a uniform spectral
value but by the distribution of a spatial autocorrela-
tion. Thereafter, the classification is not based on pixels
but on objects as the minimum information unit, as
reported for QuickBird imagery (Castillejo-Gonza
´lez
et al., 2009). Since a number of studies have shown the
spatial correlation and patchiness of broad-leaved
and grass weed species (Heijting et al., 2007b; Jurado-
Expo
´sito et al., 2009), the reasoning behind the segmen-
tation process is quite straightforward: if weeds are
distributed in patches, it is likely that the classification of
objects consisting of similar adjacent pixels will improve
the weed mapping, particularly if merely the presence or
absence of weeds is the goal.
One of the keys for practical SSWM is to translate
the detected weed pressure into management zones. To
define such zones, three questions were raised (Fridgen
et al., 2004): (i) which information should be used (e.g.
the economic threshold)? (ii) how is this information
processed? and (iii) how many zones should be estab-
lished within a field to make the management feasible
for the farmer? Fuzzy clustering algorithms (Meyer
et al., 2004), development of software for automatic
assessments of homogeneous areas (Garcı
´a-Torres et al.,
2008) and segmentation techniques (Costa et al., 2007)
Weed detection for site-specific weed management 5
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
have been applied to remotely sensed data to define less
than three or four management zones, with satisfactory
results for soil management (Khosla et al., 2007; Song
et al., 2009) and for weed management in sunflower
(Pen
˜a-Barraga
´net al., 2010) at the research level.
However, it would be beneficial to develop further
information for the delineation of management zones
for herbicide application in a larger number of crops, to
offer real solutions to farmers.
In conclusion, there are opportunities for using
objects vs pixels as minimum information units for weed
classification and mapping in remote sensing, and a
cluster analysis (or other analysis) for delineating a
restricted number of management zones. An additional
task would be to assess the accuracy of zone maps based
on QuickBird or GeoEye pan-sharpened imagery, which
can offer solutions on large scales. Finally, airborne
hyperspectral sensors have the desirable spatial resolu-
tion and mission flexibility for mapping seedling and
mature weeds in crops, but operating costs and lack of
companies that provide a cost-effective product make
them unaffordable for farmers and consultants, at the
moment.
Proximal sensing: mapping and real-time
control
For site-specific control on finer spatial scales, there is
interest in monitoring weeds using digital cameras, or
spectral or optical sensor systems (non-imaging sensors)
from ground-based platforms. Brown and Noble (2005)
and Slaughter et al. (2008) reviewed the use of ground-
based non-imaging sensors to identify crops, weeds and
background soil. Non-imaging optoelectronic sensors
have been developed by research groups (Dammer &
Wartenberg, 2007; Wang et al., 2007) or are commer-
cially available (e.g. WeedSeeker
, http://www.ntechin-
dustries.com) for real-time spraying on cereals, pea or
fallows and can be adapted for crops, such as vineyards
and orchards. In these situations, all vegetation is
assumed to be weeds and the only task is to discrim-
inate between plants and soil. Thus, this review only
includes research on automatic image processing
systems.
High-resolution on-ground monitoring can be used in
both map-based and real-time SSWM. Map-based
involves two approaches: (i) an automatic weed detec-
tion system to input digital images and a subsequent
computer-based image analysis to map the weed distri-
bution and (ii) a real-time image analysis consisting of
an automatic weed detection system mounted in the
front of a tractor (or similar) analysing the images in
real-time, but spraying later using the georeferenced
map of weed distribution in a subsequent field opera-
tion. The first approach requires a digital camera and a
differential global positioning system (DGPS), and the
second one requires a digital camera, a computer-based
image analysis with a plant identification algorithm and
DGPS.
The real-time SSWM needs are for a vehicle to have
sensing, decision-making and weed control implements
for site-specific treatments. Real-time SSWM includes
two approaches: (i) a weed detection-tractor-sprayer
combination, in which variable spraying is applied
according to weed detection in a single operation
(usually used for extensive crops such as cereals) and
(ii) a small autonomous vehicle that integrates detection
and control of weeds also in a unique and simultaneous
operation (robotic weeding) (usually used for high value
crops such as tomatoes). The second approach would
aim to replace traditional large tractors with small
autonomous machines. The next sections will present the
main findings of proximal sensing for both weed
mapping and real-time SSWM.
On-ground image analysis for weed mapping
There are a number of concepts or parameters that are
important in automatic image processing. Machine
vision is the term that refers to the capture of on-ground
images, and rule-based pattern recognition refers to the
extraction of quantitative features (Guyer et al., 1993).
Gerhards and Christensen (2003) and Gerhards and
Oebel (2006) used differential images (NIR-visible)
obtained with a set of three digital bi-spectral cameras
mounted in the front of a sprayer at a speed of
5–8 km h
)1
taking around 3000 images ha
)1
. They anal-
ysed the images in real-time, identifying characteristic
shape features of crop and weeds, but the spraying was
conducted in a later field operation. They grouped weed
species according to their sensitivity to herbicides and
achieved a fast variation of herbicide mixture according
to weed species distribution, reaching a reduction of
herbicide from up to 20% to 90%, and from 5% to 81%
for grass and broad-leaved weed herbicides in winter
cereals and rape, maize and sugar beet. Sun et al. (2010)
have developed an automatic, centimetre-level accuracy
mapping system for real-time mapping of every plant of
vegetable crops, such as tomato, during transplanting.
The crop map generated is suitable for subsequent
between-rows and within-row mechanical weed control.
Several ground-based systems, such as tripod (e.g. a
digital camera at 0.45–0.50 m above the soil surface
sampling continuous transects) Berge et al. (2008), and
tractors (e.g. digital cameras at 1.7 m above the ground
level) at a speed of 2.25 m s
)1
(Hague et al., 2006) or
4kmh
)1
(Tellaeche et al., 2008), have been tested for
taking the digital RGB (red-green-blue) imagery for
6FLo
´pez-Granados
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
subsequent image classification. Eddy et al. (2008) used
hyperspectral imagery acquired at ground level at 1 m
target distance (1.25 mm pixel and 400–1000 nm spectral
range at 10 nm intervals) and compared a new hybrid
segmentation-artificial neural network method to a
standard maximum-likelihood classification, for discrim-
ination of redroot pigweed and wild oat in oilseed rape,
pea and wheat, in both single date and multitemporal
data. They discriminated weed species and crops with
overall accuracies from 84% to 92% for multitemporal
classifications and using the new algorithm, which implies
improvements up to 31% over the standard weed control
method.
At this point, it is essential to elucidate the importance
of always using the same concept to compare different
investigations. Thus, Burgos-Artizzu et al. (2009)
attempted to clarify the controversy regarding the pos-
sible use of weed biomass, weed density, weed pressure (a
visual estimate of the percentage that weeds contribute to
the total volume of crops and weeds in a given area,
considering the volume estimates as both height and
surface area simultaneously covered by crop and weed),
weed coverage, and the relative leaf area of weeds (Ôweed
coverÕ), as some of the different parameters to be used in
on-ground image analysis. They suggested that an image
processing system that estimates the relative weed leaf
area per image is much better adapted to machine vision
than to the other weed abundance measures, because
correlation coefficients with current data were 80%.
One of the on-ground automated image processing
for crop and weeds detection has been the segmentation
of images for object-oriented analysis by converting
every (usually) RGB image to a binary black & white
image. The first step is to obtain an image where white
pixels represent vegetal cover (crops and weeds) and
black ones the soil. In the second step, the zones
corresponding to crops are identified and eliminated,
and finally, the location of weeds is obtained (Burgos-
Artizzu et al., 2009). The leaf shape, colour and texture
for each individual object in the image have been
successfully used to distinguish several weed species and
volunteer potatoes at the seedling stage in maize,
soyabean, wheat and sugar beet (Guyer et al., 1986,
1993; Zhang & Chaisattapagon, 1995; Nieuwenhuizen
et al., 2007). Tellaeche et al. (2008) reported the results
of weed detection in maize through a two-step process:
(i) image segmentation in cells of 8500 and 1660 pixels
for cells in the bottom and upper part of the images
because of the perspective of them, and (ii) a decision-
making process to determine whether or not a cell is to
be sprayed. The tractor speed was 4 km h
)1
, which
implied that 12 m was covered in 11 s. Other works have
applied fuzzy logic algorithms and artificial neural
network classifiers to discriminate young weed species
in maize at the two- to five-leaf stage (Yang et al., 2002)
and sunflower at the four-leaf stage (Kavdir, 2004).
They analysed hundreds of images and there could be
enormous differences between images. These differences
are the key to understanding the complexity of a robust
analysis of every image, and they are directly related to
the performance of the image analysis process. There-
fore, a classifier of soil, weeds and crops has to be
powerful and flexible in a number of field situations.
Burgos-Artizzu et al. (2009) discriminated barley, soil,
A. sterilis and Papaver rhoeas L. (poppy) in outdoor
field images under varying light conditions, soil back-
ground texture and crop conditions over 4 years by
means of a Case-Based Reasoning system.
Most of the research reviewed has concluded that
subsequent investigations should address on-line weed
identification for real-time herbicide application, which
is the main issue addressed in the next section.
Decision-making and real-time robotic weed control
To date, very few complete real-time robotic site-specific
weeders have been tested under field conditions. Lee
et al. (1999) correctly identified 73% of the tomatoes
and 69% of the weeds (thirteen species) testing their
robot for real-time spraying of in-row grass and broad-
leaved weeds. The prototype travelled at a speed of
1.2 km h
)1
, with 58.10 and 37.44 ms for the execution
time to find the tomato and weed locations, and to send
tomato and weed locations to the spray controller
respectively. Slaughter et al. (2008) reviewed real-time
robotic weeding, including methods for improving the
accuracy of detecting weeds under varying natural
illumination conditions, as well as methods for mechan-
ical, thermal, herbicide and electrical weed control in
several crops. They concluded that vibrations, dust and
other issues associated with real-time machine vision
systems need to be overcome.
An additional challenge of SSWM weed control
based on on-ground image analysis for weed mapping or
for real-time is to determine the correct herbicide
application rate according to the weed classification.
Some authors have developed or applied decision
support algorithms, such as WeedSOFT and Decision
Algorithm for Patch Spraying (DAPS), to estimate the
optimal herbicide rate. They introduced parameters such
as relative competitiveness and the dose–response curves
of several herbicides in field conditions (Christensen
et al., 2003; Neeser et al., 2004; Rider et al., 2006). More
recently, Christensen et al. (2009) reviewed the current
status of precision sprayers, including those with direct
injection and Drop On Demand Application Systems.
There are commercially available digital video cameras
that view the crop ahead of the tractor for real-time
Weed detection for site-specific weed management 7
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
spraying on different arable and horticultural crops (e.g.
http://www.garford.com).
Limitations and opportunities of proximal
sensing
On-ground image processing procedures can be consid-
ered similar for mapping and for real-time applications.
The main difference is that in real-time robotic weeding,
image analyses, decision-making and action mechanisms
(actuators) to open the sprayers, adjust the herbicide
rate, and control the actuation, must be conducted in
one operation. The reactions of autonomous vehicles in
changing contexts for real-time SSWM are highly
dependent on the nature and reliability of the decision-
making process and how this process is used. Because of
the very short time between detection and action, the
image or sensor processing time must be greatly reduced,
necessitating the avoidance of computationally intensive
steps. Similarly, the decision-making system must be
fast, robust, flexible and simple. The main problems to
solve are related to: (i) agronomic situations, e.g. each
crop field presents enormous differences in weed species
and abundance, and crop plants can be confused as or
hidden by weeds, or vice versa, and (ii) automation
technology, e.g. robots, must be successful for emer-
gency stops, abrupt terrains or static (stones) and
dynamic (animals) obstacles. The travel speed of the
autonomous vehicle is also highly influential, because it
must be cost-effective and it is directly related to the
weed identification and position error of weed or patch
detection. Although machine vision and real-time kine-
matic (RTK) GPS guidance systems for SSWM by
between-rows mechanical cultivators are commercially
available, additional research is needed to combine
existing algorithms for herbicide spraying and to test the
performance of these technologies under a wide range of
agricultural situations.
Final comments: directions for further
research
Ferna
´ndez-Quintanilla et al. (2008) point out that
interactions between weed science and robotic and
information technologies are important, if we are to
realise the potential herbicide savings that a spatial
distribution of weeds offers. They also note that SSWM
may provide an answer to new European regulations
regarding pesticide use. Recent articles have reviewed a
number of topics relevant to the adoption and future
perspectives of SSWM: (i) the status of the farm (size,
kind of crops), farmers (education, age, interest in
new technologies) and lack of compatibility between
machines (Reichardt & Ju
¨rgens, 2009), (ii) site-specific
weed control technologies (Christensen et al., 2009),
(iii) SSWM with one or several herbicides (Wiles, 2009),
and (iv) robotic weeding (Slaughter et al., 2008). Most
of the articles reviewed were not conducted by interdis-
ciplinary groups. This can be disadvantageous when
comparing the economic feasibility of site-specific weed
management with conventional systems. To overcome
the current problems, European research projects (FP7-
KBBE, 2008; FP7-NMP, 2009) are raised to configure
colonies of robots for real-time management decisions.
This research expects to improve the current vision,
suspension, guidance, energy (power) safety, decision-
making and actuator systems. These projects are inter-
disciplinary and enterprises and experts from unrelated
disciplines, such as agronomy (weed science), remote
sensing, electronic engineering, computing, economy or
physics, are collaborating. This interdisciplinary strategy
is the key to achieving progress and to responding to the
heterogeneous objectives that SSWM presents. Research
needs encompass comparing the use of remote and
proximal sensing in several crops to evaluate risks and
benefits and to calculate the economic output that an
enterprise can expect when investing in this technology.
There should be an ambitious work plan that would
allow testing of the potential success widely expected
from SSWM.
Conclusions
Site-specific weed management includes a wide range of
methods for the acquisition and analysis of information.
Based on reviewed literature, the main limitations for
operational SSWM can be summarised in four points:
(i) the educational requirements of end-users for learn-
ing new technological advances and the lack of com-
patibility between current and new machinery and
between machines from different manufacturers, (ii)
the high cost of the technology, (iii) the use of remote
sensing imagery that covers large-scale infestations to
create timely (early or late) and accurate weed maps, and
(iv) the use of robots that must usually work under a
wide range of changing environments.
Possible solutions to these constraints might include
the following: (i) the development and implementation
of uniform and cheaper standards, which may occur if
competition between companies increases, (ii) offering
an advisory service that provides technical support,
agronomic knowledge and specific training courses, (iii)
more research on UAV and high resolution satellite
imagery, and (iv) enabling the development of new
prototypes or improving the current prototypes of
robotic weeding into commercial products.
In conclusion, these major milestones could result
in both an accurate and low-cost sensor for detecting
8FLo
´pez-Granados
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
or mapping weeds in an autonomous and safe
robotic vehicle. This vehicle would have powerful
decision-making capabilities and would be equipped
with individual spray nozzles for SSWM and ⁄or using
high resolution satellite or UVA images in which a high
performance algorithm could be incorporated into the
image analysis for timely and accurate weed and crop
mapping.
Acknowledgements
This work was partially financed by the Spanish Minister
of Science and Innovation by project AGL2008-04670-
CO3-03 (FEDER). The author thanks Ms Amparo
Torre for her very helpful assistance in typing the
references and Dr Luis Garcı
´a-Torres and Dr Montserrat
Jurado-Expo
´sito for their valuable suggestions.
References
B
ACKES
M&J
ACOBI
J (2006) Classification of weed patches
in QuickBird images: verification by ground truth data.
In: EARSel (European Association of Remote Sensing
Laboratories) eProceedings, EARSel, Warsaw, Poland.
Available at: http://www.eproceedings.org/static/vol05_2/
05_2_backes1.html.
B
ARROSO
J, F
ERNA
´NDEZ
-Q
UINTANILLA
C, R
UIZ
D, H
ERNAIZ
P
&R
EW
LJ (2004) Spatial stability of Avena sterilis ssp.
ludoviciana populations under annual applications of low
rates of imazamethabenz. Weed Research 44,178–186.
B
ERGE
TW, A
ASTVEIT
AH & F
YKSE
H (2008) Evaluation of an
algorithm for automatic detection of broad-leaved weeds in
spring cereals. Precision Agriculture 9,391–405.
B
LASCHKE
T (2010) Object based image analysis for remote
sensing. Journal of Photogrammetry and Remote Sensing
65,2–16.
B
ROWN
RB & N
OBLE
SD (2005) Site-specific Weed
management: sensing requirements – what do we need to
see? Weed Science 53,252–258.
B
ROWN
RB & S
TECKLER
J-PGA (1995) Prescription maps for
spatially variable herbicide application in no-till corn.
Transactions of the ASAE 38,1659–1666.
B
URGOS
-A
RTIZZU
XP, R
IBEIRO
A, T
ELLAECHE
A, P
AJARES
G&
F
ERNA
´NDEZ
-Q
UINTANILLA
C (2009) Improving weed pressure
assessment using digital images from an experience-based
reasoning approach. Computers and Electronics in
Agriculture 65,176–185.
C
ASTILLEJO
-G
ONZA
´LEZ
IL, L
O
´PEZ
-G
RANADOS
F, G
ARCI
´A
-
F
ERRER
Aet al. (2009) Object and pixel-based classification
for mapping crops and their agri-environmental associated
measures in QuickBird images. Computers and Electronics in
Agriculture 68,207–215.
C
HRISTENSEN
S, H
EISEL
T, W
ALTER
AM & G
RAGLIA
E (2003)
A decision algorithm for patch spraying. Weed Research
43,276–284.
C
HRISTENSEN
S, S
ØGAARD
HT, K
UDSK
Pet al. (2009) Site-
specific weed control technologies. Weed Research 49,233–
241.
C
OSTA
JP, M
ICHELET
F, G
ERMAIN
C, L
AVIELLE
O&G
RENIER
G
(2007) Delineation of vine parcels by segmentation of
high resolution remote sensed images. Precision Agriculture
8,95–110.
D
AMMER
KH & W
ARTENBERG
G (2007) Sensor based weed
detection and application of variable herbicide rates in
real time. Crop Protection 26,270–277.
D
E
C
ASTRO
AI, J
URADO
-E
XPO
´SITO
M, P
EN
˜A
-B
ARRAGA
´N
JM,
G
ARCI
´A
-T
ORRES
L&L
O
´PEZ
-G
RANADOS
F (2009)
Discriminacio
´n de malas hierbas crucı
´feras en cultivos de
invierno para su aplicacio
´n en agricultura de precisio
´n. In:
Actas XIII Congreso de la Asociacio
´n Espan
˜ola de Tele-
deteccio
´n(ed. S M
ONTESINOS
&LF
ERNA
´NDEZ
), 61–65.
Compobell, Calatayud, S.L., Spain.
E
DDY
PR, S
MITH
AM, H
ILL
BD, P
EDDLE
DR, C
OBURN
CA &
B
LACKSHAW
RE (2008) Hybrid segmentation-artificial neural
network classification of high resolution hyperspectral
imagery for site-specific herbicide management in
agriculture. Photogrammetric Engineering and Remote
Sensing 74,1249–1257.
F
ERNA
´NDEZ
-Q
UINTANILLA
C, Q
UADRANTI
M, K
UDSK
P&
B
A
´RBERI
P (2008) Which future for weed science? Weed
Research 48,297–301.
FP7-KBBE (2008) Framework Programme 7 – KBBE: Food,
Agriculture and Fisheries, and Biotechnology (Knowledge
Based Bio-Economy). FutureFarm Project (2008–2010):
Meeting the challenges of the farm of tomorrow by
integrating Farm Management Information Systems to
support real-time management decisions and compliance to
standards European Union. Available at: http://www.
futurefarm.eu/ (last accessed: 16 September 2010).
FP7-NMP (2009) Framework Programme 7 – NMP:
Nanosciences, Nanotechnologies, Materials and New
Production Technologies. RHEA Project (2010–2014):
Robot Fleets for Highly Effective Agriculture and Forestry
Management. Available at: http://www.car.upm-csic.es/
RHEA (last accessed: 16 September 2010).
F
RIDGEN
JJ, K
ITCHEN
NR, S
UDDUTH
KA, D
RUMMOND
ST,
W
IEBOLD
WJ&F
RAISSE
W (2004) Management Zone
Analyst (MZA): software for subfield Management Zone
Delineation. Agronomy Journal 96,100–108.
G
ARCI
´A
-T
ORRES
L, P
EN
˜A
-B
ARRAGA
´N
JM, L
O
´PEZ
-G
RANADOS
F,
J
URADO
-E
XPO
´SITO
M&F
ERNA
´NDEZ
-E
SCOBAR
R (2008)
Automatic assessment of agro-environmental indicators
from remotely sensed images of tree orchards and its
evaluation using olive plantations. Computers and
Electronics in Agriculture 61,179–191.
G
ERHARDS
R&C
HRISTENSEN
S (2003) Real-time weed
detection, decision making and patch spraying in maize,
sugar beet, winter wheat and winter barley. Weed Research
43,385–392.
G
ERHARDS
R&O
EBEL
H (2006) Practical experiences with a
system for site-specific weed control in arable crops using
real-time image analysis and GPS-controlled patch spraying.
Weed Research 46,185–193.
G
IBSON
KD, D
IRKS
R, M
EDLIN
CR & J
ONSTON
L (2004)
Detection of weed species in soybean using multispectral
digital images. Weed Technology 18,742–749.
G
OEL
PK, P
RASHER
SO, L
ANDRY
JA et al. (2003a) Potential of
airborne hyperspectral remote sensing to detect nitrogen
Weed detection for site-specific weed management 9
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
deficiency and weed infestation in corn. Computers and
Electronics in Agriculture 38,99–124.
G
OEL
PK, P
RASHER
SO, L
ANDRY
JA, P
ATEL
RM & V
IAU
AA
(2003b) Hyperspectral image classification to detect weed
infestations and nitrogen status in corn. American Society
of Agricultural Engineers 46,539–550.
G
O
¨KTOG
˘AN
AH, S
UKKARIEH
S, B
RYSON
M, R
ANDLE
J, L
UPTON
T&H
UNG
C (2010) A Rotary-wing unmanned air vehicle
for aquatic weed surveillance and management. Journal of
Intelligent and Robotic Systems 57,467–484.
G
RAY
JCJ, S
HAW
DR, G
ERAD
PD&B
RUCE
LM (2008) Utility
of multispectral imagery for soybean and weed species
differentiation. Weed Technology 22,713–718.
G
UTIE
´RREZ
-P
EN
˜A
PA, L
O
´PEZ
-G
RANADOS
F, P
EN
˜A
-B
ARRAGA
´N
JM, J
URADO
-E
XPO
´SITO
M&H
ERVA
´S
-M
ARTI
´NEZ
C (2008)
Logistic regression product-unit neural networks for
mapping Ridolfia segetum infestations in sunflower crop
using multitemporal remote sensed data. Computers and
Electronics in Agriculture 64,293–306.
G
UYER
DE, M
ILES
GE, S
CHREIBER
MM, M
ITCHELL
OR &
V
ANDERBILT
VC (1986) Machine vision and image
processing for plant identification. Transactions of the
ASAE 29,1500–1507.
G
UYER
DE, M
ILES
GE, G
AULTNEY
LD & S
CHREIBER
MM
(1993) Application of machine vision to shape analysis
in leaf and plant identification. Transactions of the ASAE
36,163–171.
H
AGUE
T, T
ILLET
ND & W
HEELER
H (2006) Automated crop
and weed monitoring in widely spaced cereals. Precision
Agriculture 7,21–32.
H
EIJTING
S, W
ERF
WV, S
TEIN
A&K
ROPFF
MJ (2007a) Are
weed patches stable in location? Application of an
explicitly two-dimensional methodology. Weed Research
47,381–395.
H
EIJTING
S, W
ERF
WV, K
RUIJER
W&S
TEIN
A (2007b) Testing
the spatial significance of weed patterns in arable land using
MeadÕs test. Weed Research 47,396–405.
H
ENGL
T (2006) Finding the right pixel size. Computers and
Electronics in Agriculture 32,1283–1298.
H
ERWITZ
SR, J
OHNSON
LF, D
UNAGAN
SE et al. (2004) Imaging
from an unmanned aerial vehicle: agricultural surveillance
and decision support. Computers and Electronics in
Agriculture 44,49–61.
J
URADO
-E
XPO
´SITO
M, L
O
´PEZ
-
GRANADOS
F, G
ONZA
´LEZ
-
A
NDUJAR
JL & G
ARCI
´A
-T
ORRES
L (2004) Spatial and
temporal analysis of Convolvulus arvensis L. populations
over four growing seasons. European Journal of Agronomy
21,287–296.
J
URADO
-E
XPO
´SITO
M, L
O
´PEZ
-
GRANADOS
F, G
ONZA
´LEZ
-
A
NDUJAR
JL & G
ARCI
´A
-T
ORRES
L (2005) Characterizing
population growth rate of Convolvulus arvensis L. in
wheat-sunflower no-tillage systems. Crop Science
45,2106–2112.
J
URADO
-E
XPO
´SITO
M, L
O
´PEZ
-
GRANADOS
F, P
EN
˜A
-B
ARAGA
´N
JM
&G
ARCI
´A
-T
ORRES
L (2009) A digital elevation model to aid
geostatistical mapping of weeds in sunflower crops. Agron-
omy for Sustainable Development 29,391–400.
K
ARIMI
YA, P
RASHER
SO, M
CNAIM
H, B
ONNELL
RB,
D
UTILLEUL
P&G
OEL
PK (2005) Discriminant analysis of
hyperspectral data for assessing water and nitrogen stresses
in corn. Transactions of the ASAE 48,805–813.
K
AVDIR
I (2004) Discrimination of sunflower, weed and soil by
artificial neural networks. Computers and Electronics in
Agriculture 44,153–160.
K
HOSLA
R, I
NMAN
D&W
ESTFALL
DG 2007. Site-specific
management zones: seven years of research in the irrigated
Western Great Plains of the US. In: Proceedings Precision
Agriculture 07, Skiathos, Greece (ed. JV S
TAFFORD
),
607–613. Wageningen Academic Publishers, Wageningen,
the Netherlands.
K
OGER
CH, S
HAW
DR, W
ATSON
CE & R
EDDY
KN (2003)
Detecting late-season weed infestations in soybean
(Glycine max). Weed Technology 17,696–704.
L
AMB
DW & B
ROWN
RB (2001) Remote-sensing and mapping
of weeds in crops. Journal of Agricultural Engineering
Research 78,117–125.
L
AMB
DW & W
EEDON
MM (1998) Evaluating the accuracy
of mapping weeds in fallow fields using airborne digital
imaging
(
Panicum effusum
)
in oilseed rape stubble. Weed
Research 38,443–451.
L
AMB
DW, W
EEDON
MM & R
EW
LJ (1999) Evaluating the
accuracy of mapping weeds in seedling crops using airborne
digital imaging: Avena spp in seedling triticale. Weed
Research 39,481–492.
L
EE
WS, S
LAUGHTER
DC & G
ILES
DK (1999) Robotic
weed control system for tomatoes. Precision Agriculture
1,95–113.
L
O
´PEZ
-G
RANADOS
F, J
URADO
-E
XPO
´SITO
M, P
EN
˜A
-B
ARRAGA
´N
JM&G
ARCI
´A
-
TORRES
L (2006) Using remote sensing for
identification of late-season grass weed patches in wheat.
Weed Science 54,346–353.
L
O
´PEZ
-G
RANADOS
F, P
EN
˜A
-B
ARRAGA
´N
JM, J
URADO
-E
XPO
´SITO
Met al. (2008) Multispectral classification of grass weeds
and wheat (Triticum durum) using linear and nonparametric
functional discriminant analysis and neural networks.
Weed Research 48,28–37.
M
ARTI
´N
P, B
ARRETO
L, R
IAN
˜O
D, F
ERNA
´NDEZ
-Q
UINTANILLA
C,
V
AUGHAN
P&S
ANTIS
A (2009) Cartografı
´a de malas
hierbasen cultivos de maı
´z mediante ima
´genes hiperespec-
trales aeroportadas (AHS). In: Actas XIII Congreso de la
Sociedad Espan
˜ola de Teledeteccio
´n(eds S M
ONTESINOS
&
LF
ERNA
´NDEZ
), 41–44. Compobell, Calatayud, S.L.,
Spain.
M
EDLIN
CR, S
HAW
DR, G
ERARD
PD&L
A
M
ASTUS
FE (2000)
Using remote sensing to detect weed infestations in Glycine
max.Weed Science 48,393–398.
M
EYER
GE, N
ETO
JC, J
ONES
DD & H
INDMAN
TW (2004)
Intensified fuzzy clusters for classifying plant, soil, and
residue regions of interest from colour images. Computers
and Electronics in Agriculture 42,161–180.
M
ORAN
MS, I
NOUE
Y&B
ARNES
EM (1997) Opportunities and
limitations for image-based remote sensing in precision crop
management. Remote Sensing of Environment 61,319–346.
N
EESER
C, D
ILLE
JA, K
RISHNAN
Get al. (2004) WeedSOFT: a
weed management decision support system. Weed Science
52,115–122.
N
IEUWENHUIZEN
AT, T
ANG
L, H
OFSTEE
JW, M
U
¨LLER
J&
H
ENTEN VAN
EJ (2007) Colour based detection of volunteer
potatoes as weeds in sugar beet fields using machine
vision. Precision Agriculture 8,267–278.
P
EN
˜A
-B
ARRAGA
´N
JM, L
O
´PEZ
-
GRANADOS
F, J
URADO
-E
XPO
´SITO
M&G
ARCI
´A
-T
ORRES
L (2006) Spectral discrimination of
10 FLo
´pez-Granados
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
Ridolfia segetum and sunflower as affected by phenological
stage. Weed Research 46,10–21.
P
EN
˜A
-B
ARRAGA
´N
JM, L
O
´PEZ
-
GRANADOS
F, J
URADO
-E
XPO
´SITO
M&G
ARCI
´A
-T
ORRES
L (2007) Mapping Ridolfia segetum
patches in sunflower crop using remote sensing. Weed
Research 47,164–172.
P
EN
˜A
-B
ARRAGA
´N
JM, L
O
´PEZ
-
GRANADOS
F, J
URADO
-E
XPO
´SITO
M&G
ARCI
´A
-T
ORRES
L (2010) Sunflower yield as related to
weed infestation, land elevation and multitemporal remote
sensing imagery. Precision Agriculture 11,568–585.
R
EICHARDT
M&J
U
¨RGENS
C (2009) Adoption and future
perspective of precision farming in Germany: results of
several surveys among different agricultural target groups.
Precision Agriculture 10,73–94.
R
IDER
TW, V
OGEL
JW, D
ILLE
JA, D
HUYVETTER
KC & K
ASTENS
TL (2006) An economic evaluation of site-specific herbicide
application. Precision Agriculture 7,379–392.
S
CHMALE
DG, D
INGUS
BR & R
EINHOLTZ
C (2008) Develop-
ment and application of an autonomous unmanned aerial
vehicle for precise aerobiological sampling above agricul-
tural fields. Journal of Field Robotics 25,133–147.
S
HAW
DR (2005) Translation of remote sensing data into weed
management decisions. Weed Science 53,264–273.
S
LAUGHTER
DC, G
ILES
DK & D
OWNEY
D (2008) Autonomous
robotic weed control systems: a review. Computers and
Electronics in Agriculture 61,63–78.
S
ONG
X, W
ANG
J, H
UANG
W, L
IU
L, Y
AN
G&P
U
R (2009)
The delineation of agricultural management zones with high
resolution remotely sensed data. Precision Agriculture 10,
471–487.
S
RINIVASAN
A (2006) Precision Agriculture: an Overview. In:
Handbook of Precision Agriculture: Principles and
Applications (ed. A S
RINIVASAN
), 3–18. Food Products Press,
The Haworth Press. New York.
S
UN
H, S
LAUGHTER
DC, P
E
´REZ
-R
UIZ
M, G
LIEVER
C,
U
PADHYAYA
SK & S
MITH
RF (2010) RTK GPS mapping of
transplanted row crops. Computers and Electronics in
Agriculture 71,32–37.
T
ELLAECHE
A, B
URGOS
-A
RTIZZU
XP, P
AJARES
G&R
IBEIRO
A
(2008) A vision-based method for weeds identification
through the Bayesian decision theory. Pattern Recognition
41,521–530.
T
HORP
KR & T
IAN
LF (2004) A review on remote sensing of
weeds in agriculture. Precision Agriculture 5,477–508.
W
ANG
N, Z
HANG
N, W
EI
J, S
TOLL
Q&P
ETERSON
DE (2007) A
real-time, embedded, weed-detection system for use in wheat
fields. Biosystems Engineering 98,276–285.
W
ILES
LJ (2009) Beyond patch spraying: site-specific weed
management with several herbicides. Precision Agriculture
10,277–290.
Y
ANG
CC, P
RASHER
SO, L
ANDRY
L&R
AMASWAMY
HS (2002)
Development of Neural Network for weed recognition in
corn fields. Transactions of the ASAE 45,859–864.
Z
HANG
NQ & C
HAISATTAPAGON
C (1995) Effective criteria
for weed identification in wheat fields using machine
vision. Transactions of the ASAE 38,965–974.
Weed detection for site-specific weed management 11
2010 CSIC
Weed Research 2010 European Weed Research Society Weed Research 51, 1–11
