chapter 5 image processing method for crop status...
TRANSCRIPT
Chapter 5
Image Processing Method for Crop
Status Management
This chapter discusses the concept of image processing algorithms and its implemen-
tation for crop status management issues for sugarcane crop. Primarily there are three
issues:
1. Measurement of growth
2. Estimation of the chlorophyll content and
3. Disease severity of the leaf
Section 5.1 discusses the background of sugarcane crop and further sections would
highlight technical details (Implementation and Results).
5.1 Introduction
The duration of sugarcane crop in India ranges from 10-18 months, a 12 months crop is
most common. Experimentally and by research it has been proved that for high yield
of the sugarcane, careful crop status management is essential during the germination
and tillring stages of the growth [2].
Soil quality, fertilizer and micronutrients, water and other environmental factors
such as temperature and humidity play an important role in the growth of the sug-
arcane. Compared to other crops, the cultivation of sugarcane demands more water
(150-200 times more than other crop like jawar). The application of fertilizers and
pesticides is also quite more [40]. Considering all these factors the productivity of
sugarcane crop draws attention of farmers.
53
The growth of sugarcane is measured by stalk and leaf dry weight in gram per
centimeter square but this destructive method is rarely used for growth measurement
[58]. Measurement of vertical height of the sugarcane is commonly used method in
practice to measure the growth. Traditionally, the plant geometrical parameters are
measured using tape and protector [59]. Though these measuring tools are simple, the
procedures are difficult and also results are not accurate.
The plant leaf colour is again commonly used tool to specify health status of the
plant. Chlorophyll is a green pigment found in almost all plants, which allows the
plants to obtain energy from light [20]. The loss of chlorophyll content in leaves oc-
curs due to nutrient imbalance, excessive use of pesticide, environmental changes and
ageing. Various kinds of colour plates are available for estimation of chlorophyll con-
tent of plants [60]. Chlorophyll meter (SPAD), have been developed to estimate leaf
chlorophyll content [23]. These tools are a good option to chemical analysis method
and remote sensing method used to find chlorophyll content of the plants [61].
Most of these techniques are quite accurate but they are rarely used in practice
because of high cost of SPAD meter, unavailability of remote sensing system and other
constraints.
Another issue, concerning the sugarcane agriculture management is monitoring and
controlling of the diseases. For example fungi-caused diseases in sugarcane are the most
predominant which appear as spots on the leaves. These spots prevent the vital process
of photosynthesis and hence, to a large extent it affects the growth of the plant and
consequently the yield. In case of severe infection, the leaf gets totally covered with
spots impacting a loss of yield [31].
One may suggest the use of pesticides but excessive use of pesticide against the
protection of disease, however, increases the cost of production and results in environ-
mental degradation [62]. This drawback can be removed by estimating severity of the
disease and targeting the diseased places, with the appropriate quantity and concen-
tration of pesticide.
Traditionally, the naked eye observation method is used in the production practices
to measure the severity of disease. However, the results are subjective and are not
possibly used to measure the disease precisely [26]. Grid counting method can be used
to improve the accuracy of estimation but it is a cumbersome and time consuming
procedure.
54
Figure 5.1: Details of sugarcane plant
Therefore, there is need of effective technique for growth measurement, chlorophyll
measurement and disease severity measurement which must be a cost effective and ac-
curate support for the sugarcane agriculture management. How to use image processing
algorithms to solve the issues discussed so far in next subsequent sections.
5.2 Growth Measurement
In sugarcane, a standard way of growth measurement is to measure the length between
the +1 dewlap and lower end of a main stem [63]. Where, +1 dewlap is nothing but a
fully developed leaf as shown in Figure 5.1. The dewlap acts as a hinge between the
blade and leaf sheath. Similarly, the developments of growth of sugarcane are referred
as +1 dewlap, second +2 dewlaps and so on. Undeveloped leaves, higher than the +1
leaf are called the zero dewlaps. During the early days of the sugarcane growth, the
length of the leaf sheath grows rapidly. The apparent growth of the sugarcane is the
sum of true growth along with the growth of the leaf sheath [64].
In this understanding, the length between the first dewlap and the lower end of a
main stem is considered to be an indicator of the sugarcane growth.
5.2.1 Analysis and Design
During the experimentation, three good quality eye buds of sugarcane (Sugarcane seed
type: Cultivar Co-86032) were planted in the pots (on the 1st June 2010). Soil se-
lected in pot was lumpy soil whose contents are N-164.25, P-5.38, K-69.44, Cu-0.16,
Fe-2.68, Mn-1.24, Zn-0.12. The pots were kept in open space (North-East side) so that
they got maximum sunlight. Images of the plant were captured in interval of five days
55
Figure 5.2: Sugarcane plant structures with 0 and +1 dewlap
up to 3 months and at the same time growth was measured physically using meter scale.
It was observed that growth of stem was in the straight direction and leaves were
spreading in several way alternatively. Growing structures were also observed to be
different for individual plants of sugarcane. One can easily distinguish the position
of the dewlaps in an image as the eyes are able to recognize the boundary shape and
colour features of the plants. To measure the growth using image processing algorithm
it is necessary to distinguish the distinctive boundary shape and colour features of
sugarcane plants [65]. Certain features which are considered are as follows:
1. Stem: Shape- Straight, Colour- Red, Green,
2. Leaf blade: Shape-Curve gently, Colour- Green,
3. Tip of leaf blade: Shape- Acute angle, Colour- Green,
4. Dewlap: Shape- Obtuse angle at the joint of a leaf blade and a leaf sheath,
Colour - Green.
Since the structure of each plant is different with respect to its degree of growth.
There are following three classifications of cases of plant structure:
Case 1 : Those plants which have only one leaf, is an undeveloped leaf. Such plants
are discarded from measurement procedure. Pictorial understanding of this pos-
sibility is shown in Figure 5.2(a).
Case 2 : If there is one leaf axil seen from the stem edge and the other leaf is unde-
veloped along the stem, there could be two possibilities:
56
Figure 5.3: Sugarcane plant structures with +2 dewlaps
1. If the slope of left leaf is gentler than undeveloped leaf, then left leaf is +1
dewlap and undeveloped leaf is 0 leaf.
2. If the slope of right leaf is gentler than the undeveloped leaf, then the right
leaf is +1 dewlap and undeveloped leaf is 0 leaf. Pictorial understanding of
these two possibilities is shown in Figure 5.2(b),(c).
Case 3 : If there are two leaf axils seen from the stem edge and undeveloped leaf along
the stem, then +1 dewlap is at left, +2 dewlap is at right and 0 leaf is along the
stem, as shown in Figure 5.3(a),(b).
First fully developed leaf: This is the +1 leaf of sugarcane. As per expert’s
suggestions, a leaf is termed as first fully developed leaf if angle between A and B of
plant models is ≥ 200
i.e.∠AOB = φ ≥ 200 (5.1)
Where,
A - Leaf blade boundary,
B - Stem boundary,
O- Joining point of A and B.
Thus, the concept of measuring the growth of the sugarcane plant lies between the
lower ends of the stem to +1 dewlap.
57
5.2.2 Implementation
In this research, there is a sequence of steps to be followed in order to correctly measure
the growth of the plant.
Algorithm:
1. Acquire the image
2. Resize the image and convert to proper format
3. Segment the image (Convert to digital image)
4. Based on the segmented image, measure the properties of image
5. Compute the reference height and reference factor from these properties
6. For measurement of height:
(a) Compute the properties of segmented image
(b) Smooth the image to trace object from bottom of the image
(c) Find all left and right branches
(d) Compute the angle of interaction and decide whether it is a branch
(e) When the last branch is reached, count the number of pixels from first point
to end point (last branch)
(f) Calculate growth of the plant by using reference factor calculation step.
A short description of the process and steps involved are:
1. Image acquisition : Digital camera specification- Make Nikon, 12 Mega pixels,
4X zoom. A digital camera of above specification was used to acquire images.
For proper visibility of plant and easy analysis, following terms were considered.
(a) Background: In the young stages of the plant, leaf colour is green and stem
colour shed ranges from Red to Green. In this case white background is
used for accurate segmentation.
(b) Standard reference: A 90 × 10 mm size brown strip is fixed on white back-
ground which is used as standard reference in growth measurement.
(c) Adjustable position: The distance between camera and pot is adjusted in
such a way that camera covers the full area of stem from lower end to top
end along with brown strip in an image.
58
Figure 5.4: Original image of sugarcane plant
(d) Environment: All images are captured in cool white fluorescent light.The
sample image of sugarcane plant is shown in Figure 5.4.
2. Image re-sizing : A high resolution camera was used for acquisition of colour
image size 4000×3000, the image is down sampled by factor of 10 (resized to
400× 300), for low processing time. The resizing of an image was performed by
the process of the interpolation [44].
3. RGB to grayscale conversion : Input image is first converted to the grayscale
intensity image. The formula used to covert input colour image to gray scale is
given in equation (5.2).
I = 0.3R + 0.59G+ 0.11B (5.2)
Image is captured in constant cool white fluorescent light placing the plant in
front of the white background, results into a large difference in between gray
values of two groups, object and background is as shown in Figure 5.5.
Figure 5.5: Gray scale image of sugarcane plant
59
4. Image segmentation : Image segmentation is the important step to separate
the different regions with special significance in the image, these regions should
not intersect with each other and each region should meet the consistency con-
ditions in specific regions [66]. In this experiment standard reference and plant
is separated from the white background.
• Segmentation technique: From the image statistics, maximum and minimum
gray levels are obtained and threshold value is computed by averaging of
these levels [67]. This threshold value converts the gray scale image to a
binary image. The thresholding value for sample image comes out to be 91
and the resultant binary image is as shown in Figure 5.6.
Figure 5.6: Binary image of sugarcane plant
5. Image filtering :The binary image is negated and morphological operation
(close and fill) is used to filter the image. Morphological operations are based on
relations of two sets. One is an image and the second is a small probe, called a
structuring element. The structuring element systematically traverses the image
and its relation to the image in each position is stored in the output image [47].
Closing tends to smooth sections of contour; it generally fuses narrow breaks and
the long thin gulf, eliminates small holes and fills gaps in contour. The closing
of set A by structuring element B , denoted A •B is defined as :
A •B = (A⊕B) Θ B (5.3)
Closing might result in amalgamations of disconnected components, which gen-
erated new holes. Region filling had to be adapted. The algorithms for region
filling based on set dilation, complementation, and intersections, as shown in
Equation 5.4
Xk = (Xk−1 ⊕B) ∩ AC k = 1, 2, 3, ... (5.4)
60
Where, A is a set of boundary and B is structuring element. The algorithms
terminates at iteration step k if Xk = Xk−1.
In the sample image during eroding operation lower leaf is removed from the
image because of very thin edge.
6. Computation of reference factor : The binary image includes reference ob-
ject and plant. To calculate the reference factor it is necessary to separate the
standard reference from the image. The image consists of two labeled regions
corresponding to plant and reference object. Objects labeled image is shown in
Figure 5.7.
Figure 5.7: Plant and reference object labeled image
From labeled image the area of reference object is computed. In sample labeled
image, area of reference object is 269 pixels, therefore the reference object is
separated from plant object assuming area less than 300, and greater than 30
pixels to remove unwanted pixels in the image.
• Reference factor is calculated as
Reference Height(Hr) = Max rowsize−Min rowsize (5.5)
In sample image maximum and minimum row size is 62 and 03 respectively.
Hence,
Reference Factor(Rf ) =Standard reference height
Hr
(5.6)
In sample image Hr = 59 and Standard reference height = 90mm. Thus,
Rf = 1.5254 This value will be required to calculate the actual growth of the
plant at the end.
61
Figure 5.8: Cropped image of plant
7. Computation of growth of sugarcane plant : For the sample plant shown in
Figure 5.7, pixels corresponding to the reference object are less than 300. After
morphological filtering, based on the black pixels from top to bottom, the cropped
image is as shown in Figure 5.8. This extracted area is considered as the dewlap
search region.
• Extraction of a dewlap : Dewlaps are searched on the assumption that
they appear on the stem edge in alternate position. It is difficult to search
the edge of the stem if withered leaves exist in an image, so withered leaf is
eliminated before the dewlap detection.
The boundary shape of the main stem is almost straight in each plant.
Clockwise boundary tracking is performed on the dewlap search area [68].
For the binary image, all white pixels belong to the object and all black
pixels constitute the background as shown in Figure 5.9
Figure 5.9: Boundary tracking
62
The y-coordinate of the start point was the same as the y-coordinate of the
lower end of the stem. Numbers from the start n, called boundary numbers,
and coordinates (Xp(n), Yp(n)) were recorded.
Chain coding expressing the moving direction of adjacent pixels on a bound-
ary is one way to provide for shape recognition. Since the chain pixels are
limited to -180,-135,-90,-45, 0, 45, 90 and 135 degrees.
Calculated boundary curvature which was defined by an exterior angle
formed with the present boundary pixel and pixels 20 pixels apart from
it in the backward and forward directions to increase the angular resolu-
tion. The boundary curvature φ(n) was calculated and recorded for all the
boundary pixels.
Figure 5.10 shows the curvatures at a dewlap, the tip of a leaf and an axial.
Figure 5.10: Boundary tracking in both directions
After tracing in both directions (right and left) and obtaining the bound-
ary curvature, we get exterior angle [69]. The leaf tips or axils are clearly
detected from the boundary curvature, therefore the axils that correspond
to the dewlap are detected first and then dewlaps are determined based on
the locations of the axils [70].
The boundary pixels that satisfy the following inequalities are confirmed as
candidates for axils.
Referring Figure 5.11, ∠AOB = φ ≥ 200
63
Figure 5.11: Sugarcane plant structure with +1 dewlap
Where,
A = leaf blade boundary,
B = Stem boundary and
φ=angle between A and B
If two or more pixels satisfy this inequality at one axil, the nearest one to
the low end of the stem is selected as the true axis. The detected axils are
searched by the distance from the lower end of the stem [71, 72]. The pair
(Xai, Y ai) consists of the x and y coordinates of the ith highest axial. The
objective of this process is to detect the dewlap of the +1 leaf. However the
structure of each shoot differed according to its degree of growth and facing
direction.
In sample image the structure of sugarcane plant is similar to case 3 that
described earlier in different growth structures of plant. There are two or
more axils. The boundary number of the highest axial p and that of the
second highest one q is compared. If p is smaller than q, the left side of the
shoot is searched.
Here, it is assumed that if angle φ is greater than or equal to 200 it is a fully
developed leaf.
Angle φ is calculated as below:
64
φ = aCosdpA.B
×180
π(5.7)
Where,
dp= Dot product of the vectors,
A and B = Length of the vectors and
aCos = Inverse cosine results in radians.
• Computation of growth of plant : To detect the developed leaf and its
angles, stem is traced in bottom up fashion. The top most leaf blade with
angle φ ≥ 200 and the highest distance from the lower end is the +1 dewlap
[73].
The distance of +1 dewlap from the lower end is the growth of the plant.
i.e. L = (Emax ×Rf )× Averagingfactor (5.8)
Where,
L = Growth of the plant in mm,
Emax= Maximum pixel count from lower end (as computed),
Rf= Reference factor (as computed).
Averaging factor = 0.86 (Since the pot and standard reference are not in
the same plane, we consider this factor 0.86; value is obtained after iterative
evaluation over 50 images).
For the sample image,
Emax = 47 and
Rf = 1.5254
Thus,
L = 61.65mm
The +1 dewlap detected sample image is as shown in Figure 5.12.
5.2.3 Results and Discussions
In sugarcane agriculture it is essential to monitor the growth during the first two stages
that is Germination and Till-ring for efficient fertilization and irrigation purpose. The
65
Figure 5.12: +1 dewlap detected image of plant
approximate period of these growth stages is of 120 days.
In this research the vertical growth of sugarcane plants, planted in pots is measured
by traditional method of measurement that is by using meter scale under the guidance
of sugarcane agricultural scientist. These readings are considered as standard measure-
ments (SM) for validation of results of designed algorithm for growth measurement.
For experimentation purpose, 72 sample images of sugarcane plants are tested using
the designed algorithm. Tabular representation of measurement of growth of these 72
samples is given in Appendix-C. The objective of this algorithm was to measure the
growth (i. e. the length of the +1 dewlap from the lower end of stem).
Accuracy of measurement of growth by designed algorithm is calculated by referring
the readings of standard measurements. Equation 5.9 and 5.10 gives Deviation Factor
(DF) and Accuracy (AC) as:
DF = (SM − EM)×100
SM(5.9)
AC = 100−DF (5.10)
Where,
DF - Deviation Factor,
AC - Accuracy,
SM - Standard measurement (measurement by manual method) and
EM - Experimental measurement (measurement by Algorithm)
The accuracy of measurement of growth by proposed method comes out to be 96%.
A comparison of growth measured by the proposed method and growth measured by
66
0 10 20 30 40 50 600
10
20
30
40
50
60
Growth by meter scale
Gro
wth
by
prop
osed
alg
orit
hm
Correlation between growth measurement
Plant 01 RMSE = 0.2872
Figure 5.13: Correlation between growth measured by standard and experimental mea-
surement method
meter scale is shown in Figure 5.13. The Root Mean Square Error (RMSE) of com-
puted values with respect to meter scale values is 0.2872. The overall result indicates
that there is a good agreement between growth measured by meter scale and growth
measured by proposed algorithm.
5.3 Chlorophyll Measurement
Chlorophyll is a green pigment which is the basic ingredient of the leaf of a plant [74].
It is due to the presence of chlorophyll, that the leaves are green colours in nature.
Chlorophyll absorbs certain wavelengths of light within the spectrum of visible light as
shown in Figure 5.14. It absorbs both red region (long wavelength) and the blue region
(short wavelength) of the visible light spectrum while reflect green colour wavelength
which makes the plant appear to be green [75, 76].
Plants are able to satisfy their energy requirements by absorbing light from the blue
and red parts of the spectrum. However, there is still a large spectral region between
500 to 600 nm where chlorophyll absorbs little amount of light.
67
Figure 5.14: Absorption spectra of chlorophyll
Chlorophyll measuring meters measure the optical absorption of a leaf to estimate
its chlorophyll content [77]. Chlorophyll molecules absorb in the blue and red bands,
but not the green and infra-red bands, thus meters measure the amount of absorption
at these bands to estimate the amount of chlorophyll present in the leaf.
5.3.1 Analysis and Design
From the frequency spectra of absorption of chlorophyll, it can be seen that there
is good relation between chlorophyll content of leaf and primary colours light (Red,
Green, and Blue).
The intensity of colour for each leaf disk is an arbitrary unit of intensity called
inverse integrated gray value per pixel [26]. It is proportional to the actual concentra-
tion of chlorophyll per pixel. Thus the concentration of chlorophyll per pixel can be
measured from the colour image of leaf.
The chlorophyll value of sugarcane leaf can be measured by chlorophyll meter but
it tends to be costlier [25]. It also suffers from another limitation that the readings
cannot be taken at low light intensity.
Thus there is a need of a cost effective system which can overcome the above limi-
tation.
Hence an algorithm is designed to measure the chlorophyll content of the sugarcane
leaf using RGB based image analysis. Presumably, any colour can be decomposed into
the primary colour components (RGB) and the intensity of an individual colour upto
some extent can be represented by brightness in a digital image [78]. Thus it is pos-
68
sible to develop a mathematical correlation between chlorophyll content of the leaf of
sugarcane plant and the brightness values of the primary colours.
For the purpose of verification of results, the chlorophyll value of sugarcane leaf
was measured as SPAD (Soil Plant Analysis Development) values with the help of
chlorophyll meter (CM 1000 meter manufactured by Spectrum Technology USA). The
chlorophyll meter measure transmission of Red light (650nm) at which chlorophyll
absorbs light, as well as transmission of Infrared light at 940 nm, where no absorption
occurs. The instrument calculates a SPAD value based on these two transmission values
and it is well correlated with chlorophyll content. The chlorophyll meter calculates the
SPAD values (M) by means of the Equation 5.11.
M = log[(I ′940/I940)/(I′
650/I650)] = log[(I ′940/I650)/(I′
650/I940)] (5.11)
An exponential relationship exists between the SPAD value and chlorophyll content.
The leaf chlorophyll meter output can easily be converted into the amount of leaf
chlorophyll content (µmolm−2) by the exponential equation 10M0.261, where M rep-
resents the meter output. The meter readings are considered as a standard measured
value.
Proposed image processing based method of chlorophyll measurement:
Taking into consideration, the targeted area of leaf by the Chlorophyll meter (CMY
1000) during measurement, the leaf was cut to 6× 2 cm size and the image of leaf was
captured by digital camera. From this image of leaf, mean brightness of primary colour
(R, G, and B) were recorded. The colour components were processed separately and
data were sorted based on the colour distribution as shown in Figure 5.15.
Figure 5.15: The histogram of the brightness values of the primary colours
69
The histogram represented the number of pixels occurring at each colour image
with a particular brightness value. The mean brightness values of the three- colour
components with standard deviations of a sample were obtained from the histogram
and are presented in Table 5.1.
Table 5.1: Mean brightness value with standard deviation and Coefficient of variation
Description Red Green Blue
Mean brightness value 138.02±3.60 145.88 ± 2.88 48.14± 3.82
Coefficient of variation 50 43 86.00
The small coefficients of variations for three primary colours indicated that the
mean brightness values were highly reproducible.
The mean brightness values of the three primary colours were linearly correlated to
obtain the characteristics RGB model as below
Y = aR + bG + cB (5.12)
Where, Y is predicated chlorophyll,
R, G and B are basic primary colours and
a, b and c are model parameters
In the next step these primary colours were transformed into spectral parameters
such as Hue (H), Saturation (S) and Value (V) and the chlorophyll content of the leaf
is obtained as in [78]
Y = aH + bS + cV (5.13)
Where,
Y is predicated chlorophyll,
H, S and V are spectral parameters of basic primary colours and
a, b and c are model parameters
Model parameters a, b and c are evaluated using regressive method (discussed in
detail in further section).
Thus the proposed system is designed for measurement of chlorophyll content of
sugarcane leaf that involves:
1. Operation of capturing of image
70
2. Computation of mean H, S and V values and
3. Estimation of chlorophyll content of leaf.
5.3.2 Implementation
In this research an algorithm is designed to measure the chlorophyll content of the
sugarcane leaves.
Algorithm:
1. Acquire the image of leaf
2. Preprocessing of image to convert in proper format
• Image resizing
• RGB to HSV conversion
3. Segmentation of leaf region
• Smoothing the image
4. Feature extraction
• Compute the model parameters a, b and c
• Estimate the chlorophyll value by formula
1. Acquiring the images : Following experimental set-up was made for image
acquisition:
(a) Digital camera: 12 Mega pixel, 4X zoom (Nikon make) was used for acquir-
ing the image.
(b) Background selection: Selection of background is an important step to dis-
tinguish object from background. Magenta colour is complementary to green
and therefore provides good colour contrast, making hue separation success-
ful, but sharp edge between magenta and green area results in noise during
the background separation therefore white background was used [79].
(c) Sugarcane leaf was placed flat on white background in natural light. The
optical axis of digital camera was adjusted perpendicular to the leaf plane
to shoot images.
71
2. Preprocessing of image to convert into proper format : To separate leaf
from the background, the following steps are followed.
• Re-size the image : Size of captured image is 4000×3000. The image is
down sampled by a factor of 10 (resized to 400×300), to improve the speed
of further process. The resizing of an image is performed by the process
of the interpolation [44]. This resized colour image is used for HSV space
conversion.
• RGB to HSV conversion : The hue (H) defines the colour according to
wavelength, while saturation (S) is the amount of colour. An object that
has a deep, rich tone has a high saturation and one that looks washed out
has a low saturation. The last component value (V) describes the amount
of light in the colour [80]. The main disadvantage of the RGB model is that
humans do not see colour as a mix of three primaries. Rather our vision
differentiate between hues with high or low saturation and intensity; which
makes the HSV colour model closer to the human perception than the RGB
model. In the HSV space it is easy to change the colour of an object in an
image and still retain variations in intensity and saturation such as shadows
and highlights. It is simply achieved by changing the hue component, which
would be impossible in the RGB space. This feature implied that the effects
of shading and lighting can be reduced [14].
Let r, g, b ∈ [0, 1] be the red, green, and blue coordinates, respectively,
of a colour in RGB space. Let maximum be the greatest of r, g and b and
minimum the least.
To find the hue angle, h ∈ [0, 360] for HSV space is computed as:
h =
0 if max = min(
600 × g−b
max−min+ 00
)
mod3600 if max = r
600 × b−rmax−min
+ 1200 if max = g
600 × r−g
max−min+ 2400 if max = b
(5.14)
The value of h is generally normalized to lie between 0 and 3600, and h = 0
is used when max = min (for gray values) though the hue has no geometric
meaning there, where the saturation is zero.
The values for s and v of an HSV colour are computed as:
72
Figure 5.16: Original image of sugarcane leaf
s =
0 if max = 0
max−minmax
= 1− minmax
otherwise(5.15)
v = max (5.16)
The range of HSV vector is a cube in the Cartesian coordinate system, but
since hue is really a cyclic property, with a cut at red, visualizations of these
spaces invariably involve hue circles. Cylindrical and conical depictions are
most popular, spherical depictions are also possible.
3. Segmentation of leaf region : The digital image of leaf grabbed by camera is
composed of leaf and background is shown in Figure 5.16. To obtain leaf feature,
first the leaf needed to be separated from background [81]. It is important to
choose a proper colour space for effective image segmentation.
In this experimentation, leaf segmentation is achieved by a threshold algorithm
with I1, I2 and I3 feature as follows:
I1(find(DH ≤ T1)) = 1 (5.17)
I2(find(DS ≤ T2)) = 1 (5.18)
I3(find(DV ≤ T3)) = 1 (5.19)
73
Figure 5.17: RGB and HSV colour values
Where,
DH = Difference of absolute H value (i.e. 0.33 for green colour) and actual value
of leaf for green colour,
DS = Difference of absolute S value (i.e. 0.6 for green colour) and actual value
of leaf for green colour,
DV = Difference of absolute V value (i.e. 0.1 for green colour) and actual value
of leaf for green colour,
T1, T2 and T3 = Tolerances at positive side i.e. 0.15, 0.5, and 0.5 respectively.
I1, I2 and I3 = Colour space characters.
H,S and V constants are set from Figure 5.17.
Where,
I = I1.I2.I3 (5.20)
74
Figure 5.18: Binary image of leaf
Here, I is binary image is shown in Figure 5.18, where white pixels correspond to
leaf region and black pixels correspond to background.
• Image smoothing : In second step, binary image of leaf is morphologically
processed to fill holes in the image. A hole is set of background pixels that
cannot be reached by filling in the background from the edge of the image.
Morphological process named closing (i.e. dilation followed by erosion) is
performed that tends to smooth sections of contours, it generally fuses nar-
row breaks and long thin gulf, eliminates and fills gaps in the contour, that
removes the noise pixels present in the leaf image [47]. Finally it bridges
unconnected pixels in the image.
In the third step, a new image is formed from the original image and binary
image of leaf that consist only leaf without background which includes cor-
responding R, G and B values of each white pixels from original leaf image.
Separated leaf image from background is shown in Figure 5.19. In this way
leaf is separated from background image and consumption of computing is
significantly reduced because the follow-up algorithm will only work on the
area of leaf [81].
4. Feature extraction : Feature extraction includes colour image processing steps
to measure the green colour purity to indicate the amount of chlorophyll present
in leaf. From this green image, the mean values of H, S and V space are computed.
• Computation of model parameters : Model parameters in Equation
(5.13) are determined by the least square method by using the individual
H, S and V values and chlorophyll (Y) measured by chlorophyll meter of
75
Figure 5.19: Leaf separated from the background
120 leaves. The parameters a, b and c were determined by the matrixes from
the H, S and V model as shown below
[a b c] = [ATHSV · AHSV ]
−1· AT
HSV · Y (5.21)
AHSV represents the mean spectral values and vector Y represented the
chlorophyll content of the leaves of plants determined by the chlorophyll
meter. The computed module parameters a, b and c values for sugarcane
leaves are 378.4381, 68.7543 and -216.1429 respectively.
• Estimation of chlorophyll value :Referring back to equation (5.13), the
chlorophyll of sugarcane leaf computed using HSV colour space as below:
Y = aH + bS + cV
In a sample image the spectral parameters are H, S and V are 0.2373,0.2935
and 0.234 respectively. Model parameters that have been already calculated
are a=378.438, b=68.754 and c=-216.142. Therefore the estimated chloro-
phyll value by proposed algorithm is 59.4 and chlorophyll value of the same
leaf measured by chlorophyll meter is 60.
5.3.3 Results and Discussions
The chlorophyll content of sugarcane leaves is predicted by using the HSV model.
The model parameters as determined by regressive method come out to be 378.4381,
68.7543 and -216.1429 respectively.
76
0 50 100 150 200 2500
50
100
150
200
250
Chlorophyll by meter
Chl
orop
hyll
by p
ropo
sed
algo
rith
m
Correlation between Chlorophyll measurement
RMSE = 1.9334
Figure 5.20: Correlation between actual chlorophyll and predicted chlorophyll
A comparison of estimated chlorophyll value by algorithm and actual chlorophyll
value measured by chlorophyll meter is shown in Figure 5.20. The chlorophyll meter
readings has already been found to be well correlated with leaf chlorophyll content
extracted through organic solvent method. The Root Mean Square Error (RMSE) of
estimated values of chlorophyll by algorithm with respect to chlorophyll measured by
chlorophyll meter is 1.9334, which clearly indicates that the values are almost overlap-
ping each other. Tabular representation of measurement of chlorophyll by algorithm
and by chlorophyll meter of these 120 samples is given in Appendix-D.
5.4 Disease Severity Measurement
Plant disease symptoms can be measured in various ways [82] that quantify the inten-
sity, prevalence, incidence and severity of disease.
1. Disease intensity is a general term used to describe the amount of disease present
in population.
2. Disease prevalence is the proportion of fields, countries, states etc. where the
disease is detected and reveals disease at grandeur scale than incidence.
77
3. Disease incidence is the proportion or percentage of plants diseased out of a total
number assessed.
4. Disease severity is the area (relative or absolute) of the sampling unit (leaf or
fruit) showing symptoms of disease. It is most often expressed as a percentage
or proportion.
5.4.1 Analysis and Design
In this work disease severity measure has been considered particularly for brown spot
diseased sugarcane leaves.
Brown spot is a fungal disease which causes reddish-brown to dark-brown spots
on sugarcane leaves. The spots are oval and circular in shape, often surrounded by
a yellow halve and are equally visible on both sides of the leaf [31]. The long axis of
the spot is usually parallel to the midrib. They can vary from minute dots to spots
attaining 1 cm in diameter.
The disease severity of the plant leaves is measured by counting the number of
pixels and defined as the ratio of number of pixels in diseased region and total pixels
in leaf region [83]. It is expressed as below:
Disease Severity, S =Ad
Al
(5.22)
Where,
Ad = Diseased leaf area,
Al = Total leaf area.
Here,
Ad = Pd =∑
(x,y)∈Rd
L(x, y) (5.23)
Al = Pl =∑
(x,y)∈Rl
L(x, y) (5.24)
S =Pd
Pl
(5.25)
Where,
L = Leaf image,
Pd = Total pixel in diseased region,
Pl = Total pixel in leaf region,
78
Rd = Diseased region and
Rl = Leaf region.
In sugarcane agronomy it is experimental that the length and width of leaf vary
upto 1.5 meter and 3 inches respectively. To cover the whole leaf area in the image,
the distance between camera and leaf must be adjusted proportionally. This tends to
reduce the resolution of image and cause poor segmentation of diseased part of the
leaf. Consequently to increase the accuracy of measurement of disease severity, the
leaf is cut into the pieces. Another important subject in image capturing is that the
evaporation rate of sugarcane is 150 to 200 times faster than other plants. Therefore
sugarcane leaf tends to wrinkle directly after it is separated from the stalk. So, it is
advised to capture the images quickly (without delay).
Today there are no standards to measure the extent of disease severity, [31] there-
fore experiment has been carried out to test the performance of the proposed system.
This experiment consists of drawing of standard shapes with a predefined dimension
drawn using a tool such as Paint available as accessories of Microsoft Windows Oper-
ating System is shown in Figure 5.21. The objective of this experiment is to check the
accuracy of the algorithm of measurement of disease severity. Therefore, the known
shapes such as Circle, Triangle, Square and Rectangles are given as an input to the
algorithm.
Imagination is that a green rectangle as a leaf area and small other known areas are
the disease areas. During the experimentation, these small areas and their combinations
are given as inputs to algorithm and compare the output of the algorithm as measured
area with predefined area.
Figure 5.21: Known area standard diagram
79
Physical descriptors are computed as: Percentage Accuracy (AC) and Percentage
Deviation Factor (DF) is calculated from the equation 5.9 and 5.10 as:
DF = (SM − EM)×100
SMand AC = 100−DF
The total area of green part corresponding to leaf is 61.05 cm2 and area of circle,
rectangle, square, triangle and their addition of all areas are 6.81, 0.9, 1.98, 2.49 and
12.19 cm2 respectively is shown in Figure 5.21.
To compute the disease severity more accurately the proposed system work into four
different steps that is: capture the image, compute the diseased leaf area, compute the
total leaf area and calculate the disease severity. The detailed discussion on all these
four steps is given in next subsequent subsections.
5.4.2 Implementation
To calculate the brown spot disease severity of sugarcane leaf the designed algorithm
is as follows:
Algorithm :
1. Acquire the image of diseased leaf
2. Preprocessing of image to convert in proper format:
• Resize the image
• RGB to Grayscale conversion
• RGB to HSI conversion
3. Segmentation of leaf region
4. Computation of leaf area
5. Segmentation of diseased region
• Morphological processing
6. Computation of diseased area
7. Computation of the percentage diseased severity using diseased area and leaf
area.
80
Figure 5.22: Original image of diseased leaf
1. Acquire the image of diseased leaf : The following describes the experimental
set-up that was used for image acquisition.
(a) The leaf was placed on a light panel so that the spots become more visible.
(b) Light sources of cool white fluorescent light were placed at 450 on each side
of the leaf so as to eliminate any reflections and to get even light everywhere.
(c) The camera was placed just above the leaf so as to get a better snapshot.
(d) To distinguish objects in an image an easily separable background is essential
and hence white background was used in this experimentation.
The sample image of diseased leaf is shown in Figure 5.22. This image is used
for further processing.
2. Preprocessing of image : These steps convert input sample image in more
suitable format for the segmentation.
• Re-size the image : A digital camera used for capturing image was 12
mega pixels, input colour image size is 4000 × 3000, this image is down
sampled to 256 × 256 for low processing time. The down sampling of an
image is performed by the process of the interpolation [44]. The resized
colour image is used for gray scale conversion.
• RGB to Grayscale conversion : To reduce the complexity of obtaining
leaf region and region of interest, background should be removed first there-
fore input image is converted to the grayscale intensity image. The input
81
Figure 5.23: Gray scale image of diseased leaf
colour image is converted to gray scale using equation (5.2) is as below:
I = 0.3R + 0.59G+ 0.11B
For proper contrast, the image of diseased leaf was taken on white back-
ground results into large difference in gray values of two region i.e. leaf
region and background. This gray scale image is, shown in Figure 5.23,
further used for separating leaf region from the background.
• RGB to HSI conversion : Certain aspects influencing the experimental
process are:
(a) The image quality is influenced by light changes and shadows.
(b) The vein colour is usually shallower than the leaf colour and in the early
stages of disease the green colour often fades. Therefore in the course of
disease segmentation; the phenomenon of the wrong segmentation often
exists [84, 85].
RGB colour space would not be suitable for this problem. Considering
the above facts, the input RGB colour space is transformed to HSI colour
space. HSI space is more suitable for visual characteristics of human beings.
Since the brightness component is independent of the colour component,
the colour component can be good to eliminate glare, shadow and other
light factors during colour image segmentation [80]. The sample leaf image
converted into HSI and H component of sample image is shown in Figure
5.24. The H component image is used for diseased region segmentation.
82
Figure 5.24: RGB to H space converted image of diseased leaf
3. Image segmentation : In many vision applications, it is useful to isolate the
regions of interest from the background. The level to which the subdivision is
carried out depends on the problem being solved. That is, segmentation should
stop when the objects of interest or the regions of interest in an application have
been isolated. Thresholding techniques for segmentation were used [42].
Thresholding often provides an easy and convenient way to perform segmenta-
tion on the basis of the different intensities or colours in the foreground and
background regions of an image.
In this section two different thresholding techniques are implemented to obtain
total leaf region pixels and pixels of diseased region of leaf. Brief understandings
of these techniques are as:
• Leaf region segmentation : To count the total pixels in leaf region they
must be separated from the background. The reflectance of the sugarcane
leaf is stronger than its background in grayscale images. Therefore the
following threshold function is employed to segment the leaf from the back-
ground.
83
gi(x, y) =
0 fi(x, y) ≤ Ti
fi(x, y) fi(x, y) ≥ Ti
(5.26)
Where, gi(x, y) is the segmented gray level at pixel (x, y)
fi(x, y) is the original gray level at pixel (x, y)
Ti is threshold value and i represents the red, green and blue components.
The use of histogram to separate the region is a common practice. Histogram
of gray scale image is as shown in Figure 5.25. It can be seen that it
has bimodal characteristics. The left wave crest corresponds to the leaf
region with lower gray value. There are splits between the background
region and its corresponding wave crest. It can also be seen that there is
a large difference between the gray value of leaf and background (distance
between wave crest is more than 100 in case of sample image). Thus simple
thresholding technique (using a single threshold value) would not give a
good result [83].
Figure 5.25: Histogram of grayscale image of diseased leaf
Otsu thresholding technique is used to separate the leaf region form the
background [52, 53]. Otsu technique is an overall automatic threshold se-
84
Figure 5.26: Binary image of diseased leaf
lection method, which calculates co- variance between the two groups i.e.
between leaf region and background. Largest value of co- variance is selected
as threshold value. For the sample image threshold value comes out to be
0.4902 and resultant binary image for the same is shown in Figure 5.26. The
total pixels corresponding to the leaf region can be easily counted from this
binary image.
4. Computation of leaf area : To count the total white pixels in leaf region ,
the binary image is scanned in top-down format. The number of white pixels in
the white non-zero region is counted using equation (5.14), Pl =∑
(x,y)∈Rl
L(x, y) .
The leaf pixels count for the sample image comes out to be 30913.
5. Disease region segmentation : The accurate segmentation of disease region
determines the success or failure of the experiment. As the diseases characteris-
tics varies the boundaries between diseased and healthy part of leaf also varies
which results in a weak edge.
Another problem is the presence of boundary objects. The boundary objects are
not complete and thus may hinder the process of accurate segmentation therefore
the boundary objects have to be removed. Hence triangle thresholding method
is used in this experiment [86].
85
Figure 5.27: Triangle method to select threshold value
Histogram of sample image is as shown in Figure 5.27. To select the thresholding
value, triangle is constructed by drawing a line between the maximum peak of
the histogram at brightness bmax and the lowest value bmin in the image. The
distance ‘d′ between the line and the histogram h(b) is computed for all values
of ‘b′ from ‘b′min to ‘b′max. The maximum value of distance ‘d′ is selected as the
threshold value.
In sample image, the coordinates of point ‘b′min are (13, 2), ‘b′max (175, 3220)
and the threshold value is 0.0939. Using the resultant threshold, ‘H’ component
of the HSI image is thresholded to convert the Hue image into binary image as
shown in Figure 5.28.
Further, this binary image is morphologically processed to fill holes in the binary.
A hole is a set of background pixels that cannot be reached by filling in the
background from the edge of the image. Thus the resultant image is one that
would not have noise pixels at boundary, as shown in Figure 5.29.
6. Computation of diseased area : For the sample image shown in Figure 5.19
the pixel count of diseased region image (Rd) is given by equation (5.15), Pd =∑
(x,y)∈Rd
L(x, y) and is come out to be 2730.
86
Figure 5.28: Binary image of diseased leaf
7. Computation of the percentage diseased severity using diseased area
and leaf area : Disease severity is calculated by using equation (5.20) and
always expressed in percentage as:
S =
∑
(x,y)∈Rd
L(x, y)
∑
(x,y)∈Rl
L(x, y)
%S =2730
30913× 100
%S = 8.8312
The measured disease severity is expressed in a more quantitative way by using
disease severity scale developed by agricultural scientist ‘Horsfall and Heuberger’
Figure 5.29: Filtered image of diseased leaf
87
[82]. They developed early interval scale, which is still widely perceived as a way
to save time in disease assessment. A score of disease in plants is calculated and
rated on the scale. Thus according to Table 5.2 disease severity of sample image
is of category 1.
Table 5.2: Disease severity scale developed by Horsfall and Heuberger
Category Severity
0 Not infected
1 1− 25% leaf area infected
2 26− 50% leaf area infected
3 51− 73% leaf area infected
4 >75% leaf area infected
5.4.3 Results and Discussions
In agriculture research laboratory graphical method is used to measure the disease
severity of those leaves whose veins structure is individual to carry out water and nu-
trients to and from the leaves (e.g. Soybean, Betel). In these cases it is easy to cut the
diseased part of the leaf and to mark on the graph paper.
In sugarcane leaf veins are found in bundles that is called fibro vascular bundles
that carries water and nutrients to and from the leaves. In addition to it there can
be found other fibrous tissue for aiding in shaping and mechanically strengthening the
leaves. This makes difficulty to cut the sugarcane leaf at diseased place to mark on
the graph paper. Another issue related to graphical method not suitable for sugarcane
leaf disease severity measurement is that the evaporation rate of sugarcane is 150 to
200 times faster than other plants. Therefore sugarcane leaf tends to wrinkle directly
after it is separated from the stalk.
Traditionally the naked eye observation method is used in the production practises
to measure the severity of disease. However the results are subjective and are not
accurate.
For validation of proposed algorithm standard area diagram is designed as discussed
earlier in this chapter.
88
This section describes the detailed test strategies and the results obtained. Test
cases and experiments have been built to test accuracy of the algorithm. A total of 80
diseased leaves are tested using this algorithm. A tabular representation measurement
of disease severity of these 80 samples is given in Appendix E.
As per the algorithm developed in this work, area measurements for the pre-defined
shapes are taken. A comparison is made between the two sets of values and through
this the accuracy of the software is measured.
Physical descriptors are computed by referring equation (5.9 and 5.10) as:
DF = (SM − EM)×100
SMand AC = 100−DF
Accuracy and results for individual feature is listed in Table 5.3.
Table 5.3: Comparison of measurement of features
Sr.
NoShape
Standard
Measure
(SM)
Expt.
Measure
(EM)
Deviation
(DF)
Accuracy
(AC)
1 Circle 6.81 6.6 3.1 97
2 Rectangle 0.9 0.9 0 100
3 Square 1.98 1.99 1 99
4 Triangle 2.49 2.45 1.61 98
5 Group of all above 12.19 12.42 1.87 99
6 Circle and Rectangle 7.71 7.5 2.72 97
7 Circle and Square 8.79 8.59 2.27 98
8 Circle and Triangle 9.3 9.05 2.69 97
9 Rectangle and Square 2.88 2.89 1 99
10 Rectangle and Triangle 3.39 3.35 1.18 99
11 Square and Triangle 4.47 4.44 3 97
The average accuracy of the entire feature comes out to be 98.60%, where in the
rectangle feature gives ideal results. The above data confirms the accuracy (validity)
of the system for measurement of the disease severity.
5.5 Concluding Remarks
Algorithm developed for growth measurement of sugarcane plant tested on different
conditions of the +1 dewlap. The results of designed algorithm are compared with
results of traditional method used to measure the growth of sugarcane plant (i.e.
89
protector- meter scale method) for validation.
The accuracy of measurement of the growth by algorithm is greater than 96% and
RMSE value 0.2872, indicate that there is a good agreement between growth measured
by using meter scale and proposed method. The accuracy is higher at constant light
condition, detecting the boundary curvature. We could find the axils, the leaf tips and
the dewlap. However it was difficult to determine the dewlap of the +1 leaf if the 0
leaf was almost fully developed.
In determination of chlorophyll content of sugarcane leaves using HSV colour space.
Linear mathematical HSV model is proposed to co-relate with the chlorophyll content
apart from the simple correlation analysis. For the validation of results of proposed
algorithm these results are compared with results of chlorophyll meter. The resultant
RMSE value is 1.9334, indicates that there is a good agreement between chlorophyll
measured by chlorophyll meter and proposed method.
Disease symptoms of the plant vary significantly under the different stages of the
disease. Hence the accuracy depends to some extent upon segmentation of the image.
Otsu threshold segmentation is used for leaf region segmentation but this thresholding
method is not suitable to disease region segmentation because of varying characteristics
of the disease. Triangle method of the thresholding is used to segment the disease
region. There is no any standard method to measure the disease severity of sugarcane
leaf so for the validation of the results of designed algorithm known area standard
diagram is designed. The average accuracy of the algorithm is tested using known
area standard diagram and is 98.60%. This indicates that the designed algorithm can
measure the disease severity of the leaf more accurately.
90