c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/compag

A new vision-based approach to differential

spraying in precision agriculture

Alberto Tellaeche a , Xavier P. BurgosArtizzu b , Gonzalo Pajares c,∗ ,

Angela Ribeiro b , César Fernández-Quintanilla d
a Dpto. Informática y Automática, Escuela Técnica Superior de Informática, UNED, Spain
b Instituto de Automática Industrial, CSIC, Arganda del Rey, Madrid, Spain
c Dpto. Ingenierı́a del Software e Inteligencia Artificial, Facultad Informática, Universidad Complutense, 28040 Madrid, Spain
d Centro de Ciencias Medioambientales, CSIC, Madrid, Spain

a r t i c l e i n f o a b s t r a c t

Article history: One of the objectives of precision agriculture is to minimize the volume of herbicides by
Received 7 February 2007 using site-specific weed management systems. To reach this goal, two major factors need
Received in revised form to be considered: (1) the similarity of spectral signatures, shapes, and textures between
25 July 2007 weeds and crops and (2) irregular distribution of weeds within the crop. This paper outlines
Accepted 26 July 2007 an automatic computer vision method for detecting Avena sterilis, a noxious weed grow-
ing in cereal crops, and differential spraying to control the weed. The proposed method
Keywords: determines the quantity and distribution of weeds in the crop fields and applies a decision-
Precision agriculture making strategy for selective spraying, which forms the main focus of the paper. The method
Machine vision consists of two stages: image segmentation and decision-making. The image segmentation
Weed detection process extracts cells from the image as the low-level units. The quantity and distribution
Image segmentation of weeds in the cell are mapped as area and structural based attributes, respectively. From
Multicriteria decision-making these attributes, a multicriteria decision-making approach under a fuzzy context allows us
to decide whether any given cell needs to be sprayed. The method was compared with other
existing strategies.
© 2007 Elsevier B.V. All rights reserved.

1. Introduction to detect weed seedlings by machine vision focused on

Nowadays, there is a clear preference to reducing the use of
chemicals in agriculture. Numerous technologies have been
developed to make agricultural products safer and to lower
their adverse impacts on the environment, and precision agri-
culture is a valuable component of the framework to achieve
this goal (Kropff et al., 1997; Zhang et al., 2002; Stafford,
2006).
Within that general framework, weeds can be managed
site-specifically using available geospatial and information
technologies (Gerhards and Christensen, 2006). Initial efforts
geometrical measurements such as shape factor, aspect
ratio, and length/area (Pérez et al., 2000). Later, colour images
were successfully used to detect weeds and other types
of pests (Søgaard and Olsen, 2003). Yang et al. (2003) esti-
mated weed coverage and weed patchiness based on digital
images, using a fuzzy algorithm for planning site-specific
application of herbicides. Recently, Gerhards and Oebel (2006)
used real-time differential images (NIR-VIS) obtained with
a set of three digital bispectral cameras to detect small
weed seedlings in different crops. Other approaches have
used colour indices to distinguish plant material from the

∗
Corresponding author. Tel.: +34 1 3 94 75 46; fax: +34 1 3 94 75 47.
E-mail address: pajares@fdi.ucm.es (G. Pajares).
0168-1699/$ – see front matter © 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.compag.2007.07.008
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155
background (Thorp and Tian, 2004; Ribeiro et al., 2005).
Bacher (2001) estimated weed density in a field of spring
barley by image binarization and morphology followed by the
identification of crop rows using information on distances
between rows within the crop to decide on spraying. This
process serves to make weed plants appear isolated from the
crop.
Avena sterilis L. (“winter wild oat”) is one of the most
widely distributed and abundant weeds of cereals in Spain and
other regions with Mediterranean climate, causing substan-
tial losses in these crops (Barroso et al., 2004a; Radics et al.,
2004). Although some A. sterilis plants may be found growing
singly or in small patches, the majority of them are aggre-
gated in relatively large patches (Ruiz et al., 2006), and those in
early spring, after broad-leaved weeds have been controlled by
early postemergence treatments, are practically pure stands
(Fernandez-Quintanilla, personal observation). Due to these
two features, it is relatively easy for an experienced farmer
or a technical consultant to detect patches of A. sterilis visu-
ally in the early stages of crop growth. In fields of cereals
(barley or wheat), the cereal plants grow along the furrows:
the plants growing between furrows can only be weeds. But
weeds may also grow mixed with the cereal. We sought to
detect weeds by differences in appearances: isolated plants,
small or large patches, or mixed with the crop. Three main
problems arise during detection, namely (1) irregular shapes
and different sizes of the patches, (2) spectral signature and
texture similar to those of the cereal plants, and (3) irregular
distribution of the weeds in the field. This means that methods
using only absolute sizes, shapes, textures, or spectral signa-
tures are not applicable to our experiments (Aitkenhead et al.,
2003; Onyango and Marchant, 2003; Granitto et al., 2005). The
total proportion of weeds in the field is important because
it indicates the extent of competition between weeds and
the crop (Tian et al., 1999; Ribeiro et al., 2005), but distribu-
tion has not been considered in vision-based systems to our
knowledge. Barroso et al. (2004b) studied the economic bene-
fits of using site-specific weed management systems for large
patches and numerous small patches of weeds. The dam-
age from large patches to the crops is clear; they lower the
yield substantially in the current year. When numerous small
weeds patches appear during the cereal’s growth phase, they
tend to compete with the crop aggressively. Moreover, because
weeds are more prolific in producing seeds and the seeds
persist longer in soil, a failure to control weeds creates seri-
ous problems not only in the current year but also for the
following 2–3 years (see Appendix A for details of weed den-
sity).
Hence, we propose a new method with two objectives: (1)
to determine the quantity and distribution of weeds present
in the crop and (2) to decide, based on that knowledge,
whether to undertake selective spraying to control the weeds.
The method consists of an image segmentation process
and a decision-making approach. The segmentation process
extracts cells from the image as the low-level units. The
quantity and distribution of weeds in the cell are mapped as
area and structural based attributes, respectively. From these
attributes, a multicriteria decision-making approach under a
fuzzy context allows us to decide whether any given cell needs
to be sprayed.
145
2. Materials and methods
2.1. Images
The images used for this study were those of a 1.7-ha experi-
mental field of barley on La Poveda Research Station, Arganda
del Rey, Madrid. The most common weed in the field was
A. sterilis, with densities ranging from 10 to 400 plants m−2 .
Although other weed species (Papaver rhoeas, Veronica hedaere-
folia, Lamium amplexicaule) were also present in the field, at the
time of image acquisition most of them had been killed by an
early treatment with bromoxinil and mecocrop. Images were
taken on two dates in April 2003, when the plants were at
the early tillering stage (3–5 leaves). Row spacing was 0.36 m.
Although the standard row width in the area is 0.17 m, much
wider rows are common in other semi-arid areas of North
America and Australia. Wider rows simplify weed detection.
Digital images were captured with a Sony DCR PC110E camera.
The area of each image to be processed was approximately
2.1 m × 19 m and the resolution was 1152 × 864 pixels.
The images were captured under the perspective projec-
tion, which means that areas of identical size in the field
appear under different sizes in the image, depending on their
distance from the camera. Hence, we must compute those
attributes that are independent of the perspective projec-
tion. This is achieved by establishing relative measurements
between crops and weeds instead of using absolute measure-
ments, as described in the next sections.
2.2. The proposed method
The proposed method involves two sub-processes: image seg-
mentation and decision-making. The image segmentation
process divides the image into cells and extracts those fea-
tures and attributes from each cell that make it possible to
distinguish between weeds and the crop; based on that infor-
mation, the decision-making process determines whether a
cell is to be sprayed. Such decision-making requires a set of
samples for the cells of which the decision to spray – or not –
was made in the past. Hence, we must build a knowledge base
(KB) containing sets of such samples, a stage called the off-line
process. The decision-making is carried out by computing sim-
ilarity measures between the samples stored in the KB and the
cell being processed; we call this process of decision-making
the on-line process. The image segmentation is identical for
both processes (Fig. 1).
2.3. Image segmentation: weed detection
The steps involved in the proposed image segmentation pro-
cess are acquiring and binarizing images, detecting crop
rows, partition the image into a grid of cells, and extracting
attributes from the cells.
2.3.1. Acquiring and binarizing images
As mentioned before, the images were acquired under the per-
spective projection, which implies that the crop rows tend to
converge at the vanishing point out of the field of view. The
goal of this first step was to convert the input red–green–blue
146 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155

Fig. 1 – Vision-based decision process.

(RGB) image into a binary image in which the vegetation SC , and SR structuring elements in (2) to be applied to the L, C,
(whether weeds or the crop) in the RGB image is represented and R parts, respectively.
as white points and the rest as black ones.
Various methods have been proposed for image binariza- ⎡ ⎤ ⎡ ⎤
0 0 0 1 1 0 1 1 1 0
tion (Ribeiro et al., 2005; Granitto et al., 2005; Onyango and ⎢0 0 1 1 1⎥ ⎢0 1 1 1 0⎥
⎢ ⎥ ⎢ ⎥
Marchant, 2003; Bacher, 2001; Tian and Slaughter, 1998). We ⎢ ⎥ ⎢ ⎥
selected the method described by Ribeiro et al. (2005). The seg- SL = ⎢ 0 1 1 1 0 ⎥ SC = ⎢ 0 1 1 1 0 ⎥
⎢ ⎥ ⎢ ⎥
mentation was based on the three components (R, G, and B) ⎣1 1 1 0 0⎦ ⎣0 1 1 1 0⎦
that together describe each image point. The first stage of the 1 1 0 0 0 0 1 1 1 0
segmentation transforms the original RGB image into a one- ⎡ ⎤
dimensional grey level (monochrome) image by applying the 1 1 0 0 0
following expression: ⎢1 1 1 0 0⎥
⎢ ⎥
⎢ ⎥
SR = ⎢ 0 1 1 1 0 ⎥ (2)
T(i, j) = rR(i, j) + gG(i, j) + bB(i, j) (1) ⎢ ⎥
⎣0 0 1 1 1⎦
where r, g, and b are the set of real coefficients to be selected. 0 0 0 1 1
According to Ribeiro et al. (2005), the best performance is
achieved with the following parameter values: r = −1, g = 2, and
b = −1; if T(i,j) ≤ 0, then T(i,j) = 0; if T(i,j) ≥ 255, then T(i,j) = 255. 2.3.2. Detecting crop rows and partitioning the image into
The next step was to determine the grey level threshold that a grid of cells
sets the contrast breakpoint between pixels representing veg- In the resulting binary image, all vegetation – whether weeds
etation and rest of the pixels (representing everything else: or the crop – was white and the rest was black. For detect-
shadows, stones, straw and other debris, etc.). Finally, the grey- ing the crop rows in the image, we used a Hough transform,
level image was transformed into a black-and-white image which is a well-known and robust method, especially if
to obtain a binary image. According to an earlier evaluation the rows cover the whole image (Astrand and Baerveldt,
of approaches based on different thresholds for detecting 2002; Billingsley and Schoenfisch, 1997). The Hough transform
changes in an image (Rosin and Ioannidis, 2003), the best per- obtains line equations in the normal space (Gonzalez et al.,
formance was achieved using the entropy of the histogram, 2004; Gonzalez and Woods, 2002), given by x cos  + y sin  = .
following the method described by Kapur et al. (1985). There- The Hough transform also creates an accumulator of cells
fore, we used this approach in our work. A(,) indexed by  and , where high values in a cell of
To remove spurious white pixels and to smooth the white the accumulator determine a line with the indexed param-
contours from the binarized image, we applied a morphologi- eters. Only those values of the accumulator greater than Th ,
cal opening (erosion followed by dilation) operation (Onyango a threshold set to 100 by trial and error, are allowed. Because
and Marchant, 2003; Bacher, 2001). However, because of the the orientation of the crop rows was known, we searched only
perspective projection, we had to apply three different struc- for rows with  and  consistent with this knowledge, i.e. lines
turing elements for performing the morphological opening that were near-vertical with two slopes (Fig. 2). Finally, and
operation because the central rows of the crop were near- because the crop is not usually a perfect line but has a certain
vertical whereas the rows to the left and to the right had width as well, it is likely that several accumulator cells with
different slopes. We divided the image into three strips of iden- similar indices ( and ) will have high accumulated values.
tical width: left (L), central (C), and right (R). We used the SL , This means that several lines are associated to the same crop
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155
Fig. 2 – (a) Original image, (b) segmented image displaying the line of crops and cells.
row. We merged all similar lines into a single line: given two
cells A( i , i ) and A( j ,j ), we assumed that they represented
the same crop row if | i −  j | ≤ ε1 and |i − j | ≤ ε2 where ε1 and
ε2 were set to 5 and 10, respectively. We tested the performance
of these values by trial and error.
The next step was to partition the image into a grid of cells.
This was carried out by tracing horizontal lines, i.e. lines with
equation y = kc. Due to the perspective projection, the size of
the cells decreases towards the upper part of the image and
important details coming from weeds are lost when more cells
are used. Therefore, and although we traced horizontal lines
in the whole image, we processed the cells only in the lower
part of the image. So, k = 1, . . ., n and c = 50, where n is bounded
by the height of the image. With n = 13, the details in the cells
were retained. Fig. 2(a) shows the original image captured in
a barley field and Fig. 2(b) the image after the segmentation
process.
2.3.3. Extraction attributes from cells
As mentioned in Section 1, several factors affect weed detec-
tion: irregular spatial distribution of weed patches, irregular
distribution of crop plants within a row (as a consequence
of sowing failures or gaps resulting from various accidents),
147
undefined weed shapes, similar spectral signatures and tex-
tures of the crops and weeds, and so on. Moreover, due to the
perspective projection, the cells within a single image differ
in size and shape. The attributes chosen for weed detection
must be independent of the above factors. In particular, they
must be invariant to the size and shape of cells. Therefore, we
extracted relative measurements instead of absolute ones.
For segmentation, we randomly selected a subset of 30
images from the set of 146 images available to us. From each
selected image, we selected 48 cells, i.e. a total of 1440 cells.
The number of cells classified as candidates to be sprayed was
Fa = 245 (17% of 1440). This relatively small percentage was the
focus of interest in deciding upon the differential spraying.
For the set of remaining cells (Ha = 1195), we computed the
proportion of the white area in the cells:
1  Wc
H
r= (3)
Ha Ac
c=1
where Ac is the total area of a given cell c and Wc the white
area in that cell. In this kind of cell, free of weeds, the white
area represents only crops. Each cell contains left (L) and right
148 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155
Table 1 – Number of conditions and predicates according to the distribution of patches in the cell
c1 ni = 0 c2 ni = 1 c3 1 < ni ≤ 5 c4 ni > 5
1089 155 105 91
(R) patches representing the crop areas. We found r ≈ 2/5 and
r = rl + rr where rl and rr are the corresponding ratios for the L
and R crop areas, respectively. This means that rl = rr ≈ 1/5, i.e.
each crop area covered 1/5th of the total area of a cell.
Based on the expertise criterion (see Appendix A) and tak-
ing into account the image-processing procedure, we analysed
the distribution of the patches in each cell for both sets Fa and
Ha and found the following observations:
(1) two unique patches identifying the L and R areas,
(2) two patches L and R and a number (n) of isolated patches
(small non-connected areas),
(3) a number (p) of large patches connected to L or R,
(4) L and R interconnected through a patch.
The distribution of patches (1), (2), and (3) was not mutually
exclusive; a single cell could contain patches of the three types
simultaneously. Table 1 shows the number of cells belong-
ing to each of the nine relevant categories found in the cells
we analysed. Areas in the lower half of the image below the
threshold Ta (measured in pixels), set to 3 in our experiments,
were removed beforehand because, upon observation of the
above set of cells, it was found that approximately 93% of these
small areas represented non-vegetation elements with spec-
tral signatures very similar to those of green plants that had
survived the morphological opening operation. However, the
number of cells thus removed represents only 5% of the total
set of cells. Values of the threshold greater than the above
tended to eliminate isolated weed plants, which was undesir-
able. In the upper half, this was not possible because in the
perspective projection weeds could be represented by areas
smaller than 3 pixels (see Appendix A where some details are
given about Ta ). Each category or case was identified with a
condition c1 to c9 and is assigned a separate column in Table 1.
We take a condition ci as true when the predicate (defined
below for each condition in the table) is true: ni is the number
of isolated patches in a cell and pl and pr are the number of
patches which appear as protuberances connected to the L and
R areas, respectively. The number of cells analysed fulfilling
each condition is shown in the second row. Column 1 shows
the number of cells containing only the L and R areas, i.e. with-
out isolated patches (ni = 0). Columns 2–4 show the number of
cells with ni isolated patches; the predicates are grouped based
on the number of cases found. Columns 5–8 show the number
of patches pl and pr connected with the L and R areas, respec-
tively. We found only three cells with pl and pr greater than 2
(one each with pl = 3, pr = 3, and pr = 4) and hence did not con-
sider more cases with other values for this kind of patches.
Finally, column 9 shows the number of cells with the L and R
areas interconnected.
Based only on the distribution of weed patches, under the
conditions c1 to c9 , we could not conclude definitively whether
a given cell needed to be sprayed. We therefore searched for
area-based attributes because they had been used in some
c5 pl = 1 c6 pl = 2 c7 pr = 1 c8 pr = 2 c9 L–R connected
55 44 62 44 21
earlier experiments (Granitto et al., 2005; Bacher, 2001; Pérez
et al., 2000). The area-based attributes take weed densities
into account. We observed two important cases where the
distribution of patches played an important role in the final
decision.
(1) There were cells with isolated patches distributed in the
cell with a low total density. If area measurements were
the only criterion, the decision would have been not to
spray such cells. However, as explained in Section 1 and
also in the Appendix A, weeds that may be few but dis-
tributed widely represent a risk to the current and the
following crops, and these cells must be sprayed.
(2) On the contrary, there were also patches attached to the
L and R areas which represented the crop because it had
reached a high density during its growing phase. The high
density would have led to the decision to spray—which
would have been unnecessary.
Thus, decisions based solely on area values could prove
incorrect; the method of decision-making therefore justifies
the inclusion of the following two kinds of attributes, namely
structural-based and area-based.
Accordingly, the next step was to define a procedure for
computing the values of these attributes. For each cell we
conducted the following processes.
(1) Extract the connected regions in a cell, identifying each
connected region with a unique label and its area
(Gonzalez et al., 2004).
(2) Identify the L and R labels. Regions that share the same
label are connected; those that do not are unconnected;
each cell has two regions, RL and RR , each covering 1/5th
of the cell’s total area (Ac ). (See discussion related to the Eq.
(3)). RL covers the left part of the cell and RR the right one. L
and R are the white regions inside RL and RR , respectively.
(3) Exclude the L and R regions.
(4) Compute the number ni of isolated regions; each region
has a unique label (including the L and R regions).
(5) Compute the number of patches pl and pr connected to the
L and R regions; they are the white regions with the same
labels (either L or R), once L and R are excluded (note that
connected regions have the same label).
The structural-based attributes were computed as follows.
Given a cell i, we built a nine-dimensional structural array
Si = {si1 , si2 , . . ., si9 } where each element sij is an attribute
defined as follows:
1 if cj is true
sij = j = 1, 2, . . . , 9 (4)
0 otherwise
The following are subsets of mutually exclusive elements
{si1 , si2 , si3 , si4 }, {si5 , si6 }, {si7 , si8 }, or {si1 , si9 }. This means that
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155
two elements belonging to the same subset are incompatible.
However, this does not affect the performance.
Two area-based attributes were computed and embedded
as the components of an area-vector ai ; as before, given the
cell i, this vector is ai = {ai1 , ai2 }. Let m be the total number of
connected regions in the cell i (i.e. the number of labels in the
cell) and Aij the area of the jth region. Aic is the total area of the
cell and AiL and AiR are the areas for the L and R crop regions,
respectively. AiL and AiR are computed taking into account the
number of pixels inside of the regions RL and RR as described
above in point 2).
Based on the area measurements, we computed the follow-
ing coverage values:
• crop coverage:
Cic = AiL + AiR
• weed coverage:

m Rij ≡ R(Si , Sj ) =
Ciw = Aij − Cic
j=1
• soil coverage:
Cis = Aic − (Cic + Ciw )
From Eqs. (5)–(7) we computed the components for the area-
vector ai
 Eij ≡ E(ai , aj ) = 1 −
C C C 1 + ||ai − aj ||
ai1 = iw and ai2 = iw 1 − is
Aic Cic Aic
where ai1 is defined as the weed coverage rate as described
in Tian et al. (1999) and Ribeiro et al. (2005) and ai2 can be
associated with weed pressure, also as defined in Ribeiro et al.
(2005). The area attributes are relative measurements, i.e. they
are invariant to the cell’s size (position in the image).
The following analysis allowed us to determine the range
of variability of these two values. Indeed, when the weed cov-
erage is null, i.e. there are no weeds in the cell, ai1 = 0 but if the
weeds cover the full intermediate region (i.e. Ciw = 3/5Aic ), then
ai1 = 3/5. Hence, ai1 ranges from [0,3/5]. Also, if the weed cov-
erage is null ai2 = 0. The upper limit of ai2 is achieved when Ciw
is maximum (i.e. Ciw = 3/5Aic ) and Cic minimum (i.e. Cic = 0); but
if Cic is null, it means the cell has no crops. The minimum
value we obtained for Cic was 1/10Aic . Now, assuming that
Ciw = 3/5Aic , Cis = 0.3Aic . Finally, the upper limit for ai2 can be
fixed from the Eq. (8) as 4.2. Based on these limits, we mapped
the component values of the area-vector linearly to the range
[0,1]. This was intended so that both components contribute
equitably in the computation of a similarity measurement
between two area-vectors.
The next step was to build a knowledge base (KB) containing
KB1 , representing cells that require a spray, and KB2 , repre-
senting cells that do not. Each cell j was stored in KB with
its associated attributes Sj and aj . This is the off-line pro-
cess.
149
2.4. Decision-making process
Given a new image, we apply to it the segmentation pro-
cess described in Section 2.3, extracting a set of cells i with
attributes Si and ai . The goal is to reach a decision on
each i with respect to whether it requires spraying, based
on a decision-making process that considers the similar-
ity/dissimilarity measures between each cell i and those
stored in the KB. This is the on-line process.
2.4.1. Similarity measures: benefit and cost criteria
computation
Given two structural arrays Si and Sj , we apply the string-
matching concept described in Gonzalez and Woods (2002) and
compare them component by component. Let N be the number
of elements in the structural arrays (N = 9). Let M be the num-
ber of matches between both structural arrays, where a match
(5)
occurs in the kth element if sik = sjk . A measure of similarity
between Si and Sj is defined as the ratio:
M
(9)
N
(6)
Hence Rij = 1 represents a perfect match – every element in
one array matches that in the other array – between both struc-
tural arrays (M = N) and 0 a total mismatch – the two arrays
do not match even on a single element – between Si and Sj ,
(7) i.e. M = 0. The largest value of Rij gives the best match. Given
two area vectors ai and aj , we obtain the following similarity
measure Eij :
1
(10)
(8)
where ||·||is the Euclidean norm. As the components of ai and
aj range in [0,1], the maximum dissimilarity between ai and aj
is reached when aik = 0 and ajk = 1 or vice versa, i.e. for Eij ≈ 0.59
√
and ||ai − aj || = 2. Once again, we map Eij to the range [0,1]
by applying a linear transformation taking into account these
limits. Hence, Eij is null if ai = aj (i.e. a perfect match). The
lowest value of Eij gives the best match.
From the point of view of the decision-making frame-
work, Rij /Eij are respectively the benefit/cost criteria: the
higher/lower the value, the easier it is to arrive at a decision
(Wang and Fenton, 2006).
2.4.2. Decision-making formulation
Our decision-maker uses a multicriteria decision-making
(MCDM) framework under a fuzzy context based on the work
of Wang and Fenton (2006), Gu and Zhu (2006), and Chen (2000).
Given the cell i, the MCDM is expressed as a problem with two
mutually exclusive solutions (alternatives) to the spraying of
i, namely A1 (yes) and A2 (no), one of which must be chosen.
This decision is made based on the following two crite-
ria: C1 ≡ similarity between structural arrays; C2 ≡ similarity
between area vectors. We assign a relative weight value for
each criterion: w1 for C1 and w2 for C2 . Each criterion is aver-
aged by assigning it a relative weight: w1 for C1 and w2 for C2 .
They have been fixed at 0.4 and 0.6 (w1 + w2 = 1), respectively
through a cross-validation procedure described in Section 3.1
150 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155

Table 2 – Normalized performance decision table according to the criteria and the weights
Criteria (weights)

C1 (w1 ) C2 (w2 )
 1 
Decision A1 1
RiM = Ria 1 1
E1iN = m/E1ia , m/E1ib , m/E1ic w2
 2 /M, Rib2 /M, Ric2 /M ]w1  2
A2 2
RiM = Ria /M, Rib /M, Ric /M w1 E2iN = m/Eia , m/Eib , m/Eic w2
2 2

(Duda et al., 2001). The decision about the cell i is summarized

as follows.
3. Results

(1) Compute Si and ai according to Eqs. (4) and (8), respectively. To assess the validity and the performance of the proposed
(2) Recover the set KB1 (the set comprising patterns that approach we used a set of 146 digital images, about half of
indicate the need to spray) and KB2 (the set comprising which were taken on sunny days and the rest on cloudy days.
patterns that indicate that there is no need to spray). For Because the interval between any two members of the two
each cell j in KB1 (KB2 ) compute the vectors x1i (resp. x2i ) subsets (images taken on sunny days and on cloudy days)
and y1i (resp. y2i ), was always less than 3 days, we can assume that both sam-
ples corresponded to a similar growth stage of weeds and the
xki ≡ {Ri1
k k
, Ri2 , . . . , Rijk }, crop. At this stage, in which the herbicide must be applied,
the weeds and the crop plants display similar spectral signa-
yki ≡ {Eki1 , Eki2 , . . . , Ekij }; k = 1, 2; j = 1, 2, . . . , Jk (11)
tures and textures, which is one of the problems mentioned
in the introduction. Under these circumstances, the digital
where J1 /J2 is the number of cells stored in KB1 /KB2 , respec-
images represented fundamentally different natural lighting
tively.
conditions.
(3) For each xki select the three greatest values (they represent
the benefit criterion) Riak < Rk < Rk where Rk < Rk ; for each
ib ic ij ia
yki select the three smallest values (they represent the cost 3.1. Design of a test strategy
criterion) Ekia > Ekib > Ekic where Ekij > Ekia and j = a, b, c.
(4) Build the normalized performance decision table (Table 2) The set of 146 images available was split randomly in three
where R1 , R2 , E1 , and E2 are considered triangular fuzzy subsets – B1 , B2 , and B3 – of 30, 20, and 96 images, respectively.
numbers. This justifies the choice of three values, the Each subset was segmented by applying the process described
three highest values and the three lowest values, which in Section 2.3, obtaining 48 cells for each image. Each cell j is
are fixed by thebenefit and cost described by its attributes Sj and aj , computed using Eqs. (4)
 criteria,
 respectively:
and (8), respectively.
M = max Ric 1 , R2
ic
and m = min E1ic , E2ic .
(5) Choose the best alternative. A distance mea- B1 is the subset used in Section 2.3.3, with Fa and Ha cells.
surement between two triangular fuzzy numbers The KB is loaded with KB1 = 245 (Fa ) and KB2 = 1195 (Ha ). Each
a ≡ (a1 ,a2 ,a3 ) and b ≡ (b1 ,b2 ,b3 ) is defined accord- cell is stored with its attributes. B2 is used for setting the w1
ing to the and w2 weights for the benefit and cost criteria (Section 2.4.1)
 vertex method defined in Chen (2000) as through a cross-validation procedure (Duda et al., 2001). As
1 2 2 2
d(a, b) = 3 [(a1 − b1 ) + (a2 − b2 ) + (a3 − b3 ) ]. We define before, for each image we extracted 48 cells, hence B2 provided
the ideal positive solution ≡ (1,1,1) and the ideal neg-
p+ 960 cells. Based on the expertise criterion (Appendix A), 182
ative solution p− ≡ (0,0,0). Compute the following sum of (19%) were classified as those that required spraying and the
distances: reminder (81%) as those that did not require spraying. For this
set B2 we applied the proposed decision-making process (Sec-
d+
i1
1
= d(RiM , p+ ) + d(E1iN , p+ ); d+
i2
2
= d(RiM , p+ ) + d(E2iN , p+ ) tion 2.4) using the KB and varying w1 and w2 from 0.25 to 0.75,
(12) taking into account that w1 + w2 = 1. For each combination
of weights we computed the decision error by comparing the
results of our decision-making strategy with those obtained
d−
i1
1
= d(RiM , p− ) + d(E1iN , p− ); d−
i2
2
= d(RiM , p− ) + d(E2iN , p− ) by applying the expertise criterion. We searched for the min-
(13) imum error value, which was found to be 17% with w1 = 0.39
and w2 = 0.61. Therefore, these values were then used for
The performance index for each alternative h = 1,2 is: testing B3 under the following set of five tests based on the
structural and area-based measurements.
d−
ih
+ c − d+
ih
phi = (14)
2c
• Test 1 uses only the structural array.
where c is the number of criteria (c = 2 in our approach). The • Test 2 uses only the component ai1 of the area-based vector,
best alternative h for the cell i is that with the phi value clos- i.e. weed coverage (Tian et al., 1999).
est to 1. So, if |p1i − 1| ≤ |p2i − 1| then select A1 ; otherwise, • Test 3 uses only the component ai2 of the area-based vector,
select A2 . i.e. weed pressure (Ribeiro et al., 2005).
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155 151

Fig. 3 – Labelled image with the cells “S” to be sprayed.

• Test 4 uses both the components, ai1 and ai2 , of the area-
based vector.
• Test 5 uses the structural array and both the components,
ai1 and ai2 , of the area-based vector. This is the test for
assessing the approach proposed in this paper.
Comparing the results obtained by Test 5 with those from
the rest of the tests allowed us to establish the performance
of the proposed approach. Additionally, through Tests 2 and
3, we compared the effectiveness of our approach with that of
the two strategies proposed by Tian et al. (1999) and Ribeiro et
al. (2005).
3.2. Decision-making
True No Sprayed (TN, true negative), i.e. the number of cells
correctly identified as not needing the spray.
False Sprayed (FP, false positive), i.e. the number of cells that
did not need to be sprayed but identified by the method as
those that did.
False No Sprayed (FN, false negative), i.e. the number of cells
that needed to be sprayed but identified by the method as
those that did not.
Traditionally, from these four quantities, the most used
measures for classification are those that combine the four
values (Sneath and Sokal, 1973), namely the following.
TP+TN
(1) The correct classification percentage:
 TP CCP = TP+FP+TN+FN

 TN
(2) The Yule coefficient: Yule = TP+FP + TN+FN − 1 

Given a cell i belonging to B3 , we made a decision on it (whether CCP is broadly used in computer vision tasks for assessing
to spray) by comparing its attributes with those of all j cells a classifier’s performance.
belonging to the sets KB1 and KB2 .
Test 5 uses the decision-making process described in this 3.4. Analysis of results
paper based on the fuzzy MCDM. The decision-making process
used in rest of the four tests is described below. Fig. 3 shows an image belonging to the subset B3 , which was
Test 1: ∀j, j ∈ {KB1 ,KB2 } compute mk = min{Rij }j=k , where Rij segmented and processed according to the method described
is computed according to the Eq. (9); if k ∈ KB1 the cell i is to be in this paper. The cells labelled with the symbol “S” were to be
sprayed; otherwise, it should not be treated. sprayed based on the decision-making strategy developed as
Tests 2, 3, 4: ∀j, j ∈ {KB1 ,KB2 } compute Eij according to the Eq. part of this work.
(10). Test 2 uses only ai1 and aj1 ; test 3 uses only ai2 and aj2 , and Table 3 shows the results in terms of the correct classifica-
test 4 uses both (ai1 ,aj1 ) and (ai2 ,aj2 ). Obtain Mk = max{Eij }j=k ; if tion from the five tests. We computed the CCP and Yule scores
k ∈ KB1 , the cell i is to be treated; otherwise, it should not be for the set of 96 images; since we processed 48 cells for each
sprayed. image, the number of cells tested was 4608. Larger score values
The decisions for each test were verified against those
based on human judgement (Appendix A). Thus, we could
compute a measurement for validation.
Table 3 – CCP and Yule score values for the tests
3.3. Measurements for validation (percentage of cells to be sprayed)
Test 1 Test 2 Test 3 Test 4 Test 5
The results of comparing the decisions based on expert human
CCP 73 76 79 86 92
judgement with those arrived at by deploying the different
Yule 66 69 71 82 88
tests were analysed based on the following values.
% of cells to be 37.6 32.1 30.3 24.2 20.8
True Sprayed (TP, true positive), i.e. the number of cells cor- sprayed
rectly identified as needing the spray.
152 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155

Table 4 – Categorization of cells with reference into those that need spraying and those that do not, arrived at with
different structural and area features
Structural features Area features

c1 c2 c3 c4 c5 c6 c7 c8 c9 ai1 ai2

m1 1 m2 2

Number of cells to be sprayed 959 (20.8%) 31 89 223 616 314 217 41 53 111 0.72 0.0875 0.66 0.0912
Number of cells not to be sprayed: 3649 (79.2%) 2147 1215 216 71 436 321 132 98 15 0.19 0.0769 0.15 0.8260

indicate better performance. The third row in Table 3 displays
the percentage of cells to be sprayed.
The results in Table 3 lead to the following conclusions.
(a) The best performance was achieved by Test 5.
(b) Test 5 obtained better results than Test 4; this means
that the structural measurements improved the results
obtained by using only area-based measurements, as in
Test 4. Note that we used the same decision-making pro-
cess for all tests.
(c) Test 4 performed better than Tests 2 and 3; this means
that the combination of weed coverage and weed pressure
improved the results obtained by using either criterion
separately.
(d) The worst performance was obtained by using only struc-
tural measurements, i.e. Test 1.
Table 4 displays the classification of cells – those that need
to be sprayed and those that do not – according to the condi-
tions c1 to c9 (shown in Table 1) for the spatial features and the
average values for the area features (the standard deviation is
also displayed).
From Table 4 one cannot determine clear thresholds val-
ues in order to make the decision on spraying for use in future
experiments. Nevertheless, the following inferences can be
drawn.
(1) The greatest number of cells to be sprayed fulfil c4 and
those not to be sprayed, c1 .
(2) The average area values m1 and m2 are above/below a
hypothetical threshold fixed at 0.5 for spraying and not
spraying, respectively.
The number of combinations for all features is high and
some of them do not report significant information. Neverthe-
less, we have found groups of significant combinations, which
are reported in Table 5 due to their special relevance. A distinc-
tion is made between the two categories of cells: those that
need to be sprayed and those that do not. Also displayed is the
number and percentage of cells placed in either category by
the given combination (so long as the percentage was greater
than 80). The symbols ∧ and ∨ denote the logical “and” and “or”
operators. The area-feature values ai1 and ai2 are normalized
in the range [0,1] as explained at the end of Section 2.3.3.
From Table 5, one can see that the threshold for the area
features varies with the combination of the structural cate-
gories. The following inferences can be drawn.
(a) Combinations 1 and 2: if there are no isolated patches (c1 )
along with patches adjacent to the crop (c5 to c8 ), spraying
is required only if the area-feature values are high.
(b) Combination 3: if the patches are widely dispersed (c4 ),
spraying is required even when area-feature values are
relatively small.
(c) Combination 4: if the number of isolated patches (c2 , c3 ) is
small, with large patches joining crop lines (c9 ), spraying
is required although the area-feature values are relatively
small.
(d) Combinations 5 and 6: if area-feature values are high, irre-
spective of the number of structural features, spraying is
required; if low, spraying is not required.
(e) Combination 7: if the number of isolated patches (c1 , c2 ) is
small, with patches adjacent to the crops (c5 to c8 ), spray-
ing is not required although the area-feature values are
relatively large.
One issue to be addressed concerns the weeds occluded
under our vision-based system. The weeds are occluded when
they appear mixed with the crop and there are no weeds plants
within the rows. Because of the similar spectral signatures of
weeds and crops plants, possible occlusions can be detected
by analysing high densities of crop plants in the crop L and R

Table 5 – Combination of attributes and percentages of cells classified as to be sprayed or not to be sprayed
Category Combination of attributes No. of cells % of cells

Spray 1 c1 ∧ (c6 ∨ c8 ) ∧ (ai1 > 0.7) ∧ (ai2 > 0.8) 325 95

2 c1 ∧ (c5 ∨ c7 ) ∧ (ai1 > 0.6) ∧ (ai2 > 0.6) 297 94
3 c4 ∧ (ai1 > 0.5) ∧ (ai2 > 0.4) 524 91
4 (c2 ∨ c3 ) ∧ c9 ∧ (ai1 > 0.4) ∧ (ai2 > 0.3) 198 85
5 (ai1 > 0.8) ∧ (ai2 > 0.7) 254 82

No spray 6 (ai1 < 0.2) ∧ (ai2 > 0.1) 665 90

7 (c1 ∨ c2 ) ∧ (c5 ∨ c6 ∨ c7 ∨ c8 ) ∧ (ai1 < 0.6) ∧ (ai2 < 0.5) 1835 86
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155 153

regions (see discussion in Section 2.3.3). Indeed, if this density Vision tasks and intelligent systems. The authors are grateful
tends to cover the entire crop area (AiL or AiR ), it means that to the referees for their suggestions and constructive criticism
gaps within the crop could be filled with weeds. Additionally, of the original version of this paper.
the presence of two patches (c6 or c8 ) adjacent to the crops
could also be considered a sign of occluded weeds. The above
should be accompanied by the absence of isolated patches,
Appendix A. Expertise criterion
i.e. fulfilling c1 . We identified the following two conditions as
occlusions, depending on whether the occlusion was associ- The original images were visually analysed by an expert in
ated with the left or right crop line in the cell: order to detect the presence of A. sterilis in patches of sufficient
density and distribution to be valid targets for site-specific
Right : |AiR − 15 Aic | < ε ∧ (c6 ∨ c8 ) ∧ c1 ; (a) weed management. The human visual observation was car-
ried out guided by the segmented image through the approach
Left : |AiL − 15 Aic | < ε ∧ (c5 ∨ c7 ) ∧ c1 . (b) proposed in this paper. The expert identifies the cells to be
sprayed by (1) taking into account the density and dispersion
where ε is a tolerance value set to 0.05, it implies that AiL or of weeds, (2) analysing additional factors affecting the field
AiR are considered equal to 1/5Aic so long as the difference is and the crops, and (3) visual inspection.
no greater than ε. We found 125 cells fulfilling the above two
conditions, where 79 (63%) were placed in the ‘to be sprayed’ A.1. Weed density and dispersion
category, belonging to the combinations 1 and 2 in Table 4.
When weeds appear in large patches with a low dispersion,
the expert determines visually if the weed density is above a
4. Conclusions
threshold that is considered ‘safe’ from the point of losses in
yield of the crop. If the density is above that threshold, the
We propose a new approach to detecting weeds in row crops
cells must be sprayed. According to experimental studies, A.
for selective spraying in precision agriculture. Although this
sterilis at densities above 25 panicles m−2 (5–10 plants m−2 ) can
approach has proved its value for Avena sterilis growing in
lower the yield of winter barley by 10% (Torner et al., 1991). If
wide-row cereal crops, it can be used in many other situations
the weeds appear dispersed in small or isolated patches, the
as well, e.g. maize. We have designed the method based on two
expert also uses a different threshold. According to Barroso
subprocesses: (1) segmentation to separate weeds and crops
et al. (2005), residual infestations of A. sterilis in the range of
from the rest and (2) decision-making to determine where the
1–10 panicles m−2 (0.2–4 plants m−2 ) represent a risk of yield
herbicide should be selectively applied.
loss in the current and following 2–3 years (estimated at 15%,
The segmentation is based on a combination of basic
particularly in the following years). Based on the perspective
processing techniques. The decision-making is carried out
projection of our images (Figs. 2 and 3), taking into account
by combining both structural and area-based measurements
the focal length of the camera (about ∼20 mm), and using tri-
under a fuzzy context through MCDM. Although area-based
angulation between the objects in the field and their images,
measurements have been used before, we have established
it was calculated that the cells in the bottom part of the image
that the use of structural measurements improves the results
covered an area of approximately 0.4 m2 , with 8500 pixels. The
obtained when area-based attributes are the only attributes
cells in the 13th row (number of cells processed, n = 13) cover
used. This is because the distribution of weed patches in this
an area of 8 m2 , with approximately 1660 pixels. On average,
kind of fields must be considered. The occluded weeds must
the size of cells in each row is reduced in the next row at the
be studied in greater depth in the future to increase the per-
rate of 6% in terms of the number of pixels and increased at
centage of success.
a rate of 15% in real area (m2 ). On average, a weed plant in
An important issue to be addressed in the future is the
the first row of cells is represented by approximately 12 pixels.
robustness of the proposed approach, considering that light
Hence, taking into account the reduction in the number of pix-
conditions outdoors vary a great deal. One approach to
els (∼72%) in the 13th row, this weed plant is represented by
account for such variation is to apply homomorphic filter-
3 or 4 pixels. This justifies the choice of the Ta threshold and
ing (Gonzalez et al., 2004), which separates the illumination
the removing of small areas only from the lower half of the
and the reflectance components, thereby allowing reflectance
image. Additional studies about the dispersion are reported in
alone to be considered and illumination effects to be dis-
Barroso et al. (2006).
carded. Thus, only the reflectance of weeds, crops, and soil can
be considered. Automatic learning of the weights attached to
A.2. Additional factors
the benefit and cost criteria used during the decision-making
process should also be considered in future research.
The expert has available a risk map of the field, drawn up
after taking into account the following data: stability of weed
Acknowledgements patches between different years, latent weeds, biochemical
properties of soil, yield in previous years, and weed densities,
The authors gratefully acknowledge funding from the Span- estimated visually, at harvest in the previous years. The rel-
ish Ministry of Education and Science under grant number atively high spatial stability of A. sterilis patches has received
AGL-2005-06180-C03-03. Alberto Tellaeche is with Tekniker special attention in improving the precision of weed detec-
foundation in Eibar, Gipuzkoa, Spain working in Computer tion (Barroso et al., 2004a). Various studies (Walter et al., 1997;
154 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155
Christensen and Heisel, 1998) have used stratified weed map-
ping approaches from historical weed maps (obtained with
a low resolution) to divide the field into weed zones. There-
after, these zones are assessed with a higher resolution using
real-time detection technologies.
A.3. Visual observation
Visual inspection of the stage of growth of crops and weeds
verifies the expert decision based on the above two points.
Although such expert assessment is probably reliable
enough for practical use, we have to recognize various sources
of errors in the estimations. First of all, visual estimations
of patch size and density have some degree of uncertainty.
Although weed density can be estimated more reliably in
areas close to the observer, the degree of reliability decreases
rapidly as the distance increases. Furthermore, although it
is relatively easy to detect high weed densities visually, it is
not so easy to detect low densities (∼1 plant m−2 ). This fact,
together with the fact that weed patches often have irregu-
lar shapes and poorly defined borders, may introduce some
errors in defining the perimeter of the patch. Another potential
source of error is the uncertainty in estimating losses in yield.
Depending on the weather conditions in a given year, yield
losses caused by A. sterilis may vary considerably (Torner et
al., 1991). Because of this variability due to weather, the use of
economic thresholds for weed control has not received much
practical attention in the past—a limitation that can be over-
come by using broad infestation categories. In our work with A.
sterilis, we used four categories, with infestation levels vary-
ing on a logarithmic scale (>0.1 plants m−2 , 0.1–1 plants m−2 ,
1–10 plants m−2 , and >10 plants m−2 ). This scoring system may
also contribute to alleviating two major problems inherent
in any human assessment, namely inadequate training and
the progressive reduction in the quality of assessment due
to fatigue, which justifies the use of the automatic machine-
vision system as a guide because it is free of fatigue.
references
Aitkenhead, M.J., Dalgetty, I.A., Mullins, C.E., McDonald, A.J.S.,
Strachan, N.J.C., 2003. Weed and crop discrimination using
image analysis and artificial intelligence methods. Comput.
Electron. Agric. 39, 157–171.
Astrand, B., Baerveldt, A.J., 2002. An agricultural mobile robot
with vision-based perception for mechanical weed control.
Autonomous Robots 13, 21–35.
Bacher, B. 2001. Weed density estimation from digital images in
spring barley. Ph.D. thesis, Dept. Agricultural Sciences, Section
of AgroTechnology. The Royal Veterinary and Agricultural
University, Denmark.
Barroso, J., Fernández-Quintanilla, C., Ruiz, D., Hernaiz, P., Rew,
L.J., 2004a. Spatial stability of Avena sterilis ssp. Ludoviciana
populations under annual applications of low rates of
imazamethbenz. Weed Res. 44, 178–186.
Barroso, J., Fernández-Quintanilla, C., Maxwell, B.D., Rew, L.J.,
2004b. Simulating the effects of weed spatial patterns and
resolution of mapping and spraying on economics of
site-specific management. Weed Res. 44, 460–468.
Barroso, J., Ruiz, D., Fernández-Quintanilla, C., Hernaiz, P.,
Ribeiro, A., Diaz, B., Maxwell, B.D., Rew, L.J., 2005. Comparison
of sampling methodologies for site specific management of
Avena sterilis. Weed Res. 45, 165–174.
Barroso, J., Navarrete, L., Sánchez del Arco, M.J.,
Fernández-Quintanilla, C., Lutman, P.J.W., Perry, N.H., Hull,
R.I., 2006. Dispersal of Avena fatua and Avena sterilis patches by
natural dissemination, soil tillage and combine harvesters.
Weed Res. 46, 118–128.
Billingsley, J., Schoenfisch, M., 1997. The successful development
of a vision guidance system for agriculture. Comput. Electron.
Agric. 16, 147–163.
Chen, C.T., 2000. Extensions of the TOPSIS for group
decision-making under fuzzy environment. Fuzzy Sets Syst.
114, 1–9.
Christensen S., Heisel T. 1998. Patch spraying using historical,
manual and real time monitoring of weeds in cereals.
Zeitschrift für Pflanzenkrankheiten und Pflanzenschtz,
Sonderheft XVI, pp. 257–265.
Duda, R.O., Hart, P.E., Stork, D.G., 2001. Pattern Classification. John
Wiley, New York.
Gerhards, R., Christensen, S., 2006. Site-specific weed
management. In: Srinivasan, A. (Ed.), Handbook of Precision
Agriculture. Food Products Press, New York, pp. 185–206.
Gerhards, R., Oebel, 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 Res.
46, 185–193.
Gonzalez, R.C., Woods, R.E., Eddins, S.L., 2004. Digital Image
Processing using Matlab. Prentice Hall, New York.
Gonzalez, R.C., Woods, R.E., 2002. Digital Image Processing.
Addison-Wesley, New York.
Granitto, P.M., Verdes, P.F., Ceccatto, H.A., 2005. Large-scale
investigation of weed seed identification by machine vision.
Comput. Electron. Agric. 47, 15–24.
Gu, X., Zhu, Q., 2006. Fuzzy multi-attribute decision-making
method based on eigenvector of fuzzy attribute evaluation
space. Decision Support Syst. 41, 400–410.
Kapur, J., Sahoo, P., Wong, A., 1985. A new method for gray-level
picture thresholding using the entropy of the histogram.
Comput. Vision Graphics Image Process. 29 (3), 273–
285.
Kropff, M.J., Wallinga, J., Lotz, L.A.P., 1997. Modelling for precision
weed management. In: Precision Agriculture: Spatial and
Temporal Variability of Environmental Quality. Wiley,
Chichester, pp. 182–204.
Onyango, C.M., Marchant, J.A., 2003. Segmentation of row crop
plants from weeds using colour and morphology. Comput.
Electron. Agric. 39, 141–155.
Pérez, A.J., López, F., Benlloch, J.V., Christensen, S., 2000. Colour
and shape analysis techniques for weed detection in cereal
fields. Comput. Electron. Agric. 25, 197–212.
Radics, L., Glemnitz, M., Hoffmann, J., Czimber, G., 2004.
Composition of weed floras in different agricultural
management systems within the European climatic gradient.
In: Proceedings of the 6th European Weed Research Society
(EWRS). Workshop on Physical and Cultural Weed Control,
Lillehammer, Norway, pp. 58–68.
Ribeiro, A., Fernández-Quintanilla, C., Barroso, J., Garcı́a-Alegre,
M.C., 2005. Development of an image analysis system for
estimation of weed. In: Proceedings of the 5th European Conf.
On Precision Agriculture (5ECPA), pp. 169–174.
Rosin, P.L., Ioannidis, E., 2003. Evaluation of global image
thresholding for change detection. Pattern Recognit. Lett. 24,
2345–2356.
Ruiz, D., Escribano, C., Fernández-Quintanlla, C., 2006. Assessing
the opportunity for site-specific management of Avena sterilis
in winter barley fields in Spain. Weed Res. 46 (5), 379–387.
Sneath, P., Sokal, R., 1973. Numerical Taxonomy: The Principle
and Practice of Numerical Classification. W.H. Freeman, San
Francisco.
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 0 ( 2 0 0 8 ) 144–155
Søgaard, H.T., Olsen, H.J., 2003. Determination of crop rows by
image analysis without segmentation. Comput. Electron.
Agric. 38, 141–158.
Stafford, J.V., 2006. The role of the technology in the emergence
and current status of precision agriculture. In: Srinivasan, A.
(Ed.), Handbook of Precision Agriculture. Food Products Press,
New York, pp. 19–56.
Thorp, K.R., Tian, L.F., 2004. A review on remote sensing of weeds
in agriculture. Precision Agric. 5, 477–508.
Tian, L., Reid, J.F., Hummel, J.W., 1999. Development of a precision
sprayer for site-specific weed management. Trans. Am. Soc.
Agric. Eng. (ASAE) 42 (4), 893–900.
Tian, L.F., Slaughter, D.C., 1998. Environmentally adaptive
segmentation algorithm for outdoor image segmentation.
Comput. Electron. Agric. 21, 153–168.
155
Torner, C., Gonzalez-Andujar, J.L., Fernandez-Quintanilla, C.,
1991. Wild oat (Avena sterilis) competition with winter barley:
plant density effects. Weed Res. 31, 301–308.
Walter, A.M., Christensen, S., Heisel, T., 1997. Patch spraying
using weed maps from previous years. In: Proceedings of the
10th EWRS Symposium Poznan 1997, p. 141.
Wang, W., Fenton, N., 2006. Risk and confidence analysis for fuzzy
multicriteria decision making. Knowledge Based Syst. 19,
430–437.
Yang, C.C., Prasher, S.O., Landry, J.A., Ramaswamy, H.S., 2003.
Development of an image processing system and a fuzzy
algorithm for site-specific herbicide applications. Precision
Agric. 4, 5–18.
Zhang, N., Wang, M., Wang, N., 2002. Precision agriculture: a
worldwide overview. Comput. Electron. Agric. 36, 113–132.