Site Loader
Rock Street, San Francisco

Reducing the copiousness of agricultural weeds involves heavy trust on weedkiller applications. Although these attacks have been successful in increasing farm labor efficiency and harvest productiveness, concerns sing the economic and environmental impacts of these weed control patterns and the development of weedkiller opposition have generated involvement in placing alternate weed control schemes. An machine-controlled machine vision system which can separate between harvests and weeds can be an economically executable alternate to cut down the inordinate usage of weedkillers. In machine vision engineering, the chief constituent of the system is image processing and a categorization theoretical account to acknowledge harvests and weeds. This paper deals with the usage of support vector machines ( SVM ) for categorization in a existent clip system. The developed system has been tested to find the hardiness and truth utilizing least possible characteristics. The analysis of the categorization consequences shows over 97 % truth over 224 sample images.

Increasing productiveness and upgrading plantation systems are the major concerns for speed uping agricultural development. Weedss, perceived as unwanted workss holding adaptative features which allow them to last and reproduce in cropping system, hamper agricultural development by viing with harvests for H2O, visible radiation, dirt foods and infinite. So, weed control schemes are required to prolong harvest productiveness. There are several schemes for weed control such as taking weeds manually by human laborers, mechanical cultivation and utilizing agricultural chemicals known as weedkillers. Using weedkillers is the most common method which has inauspicious impacts on environment and human wellness. It besides raises some economic concerns. In United States, entire cost of weedkillers was about $ 16 billion in 2005 [ 1 ] . Major cost ineffective and strategic job in utilizing weedkillers system is that in most of the instances, weedkillers are applied uniformly within harvest field. There can be many parts of field holding no or few weeds but weedkillers are besides applied at that place. On the other manus, human engagement in using weedkillers is clip devouring and dearly-won. Repeated usage of the same weedkiller in a field tends to advance the outgrowth of weedkiller tolerant weeds. Over 290 biotypes of weedkiller tolerant weeds have been reported in agricultural Fieldss and gardens worldwide [ 2 ] .

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

The economic system of Bangladesh is chiefly supported by agribusiness. The public presentation of this sector has an overpowering impact on poorness relief, economic development and nutrient security. The entire arable land of Bangladesh is 8.44 million hectare [ 3 ] , which is comparatively lower than the entire population. Population force per unit area continues to put a terrible load on productive capacity. For cost-efficient land usage, harvest production and quality must be maximized and the cost of weed control must be minimized. The most normally used technique for using weedkillers in Bangladesh is to spray the weedkiller solution with a backpack sprayer. This technique is considered to be inefficient and clip consuming and recommended safety steps are seldom obtained. So, a machine vision system, holding the ability to separate between harvests and weeds and set weedkillers where there are weeds, can be a fresh attack which will heighten the profitableness and lessen environmental debasement. In this attack, images will be taken from harvest field and weeds will be identified by an machine-controlled system. The chief aim of this work is to utilize support vector machines as a categorization theoretical account to sort harvests and weeds from digital images and to find whether this theoretical account can be used in real-time.SVM was chosen because of important advantages of SVM such as good generalisation public presentation, the absence of local lower limit and the thin representation of solution [ 4 ] .

Much research has investigated assorted schemes to happen out a robust weed control system. A few real-time field systems have been developed. The exposure detector based works sensing systems developed by Shearer and Jones ( 1991 ) and Hanks ( 1996 ) were able to observe all the green workss and spray merely the workss. Islam et Al. ( 2005 ) used PDA as processing device and step Weed Coverage Rate ( WCR ) to know apart between narrow and wide foliages. Ahmad I. et Al. ( 2007 ) developed an algorithm to sort images into wide and narrow category based on Histogram Maxima with threshold for selective weedkiller application with an truth of 95 % . Ghazaliet Al. ( 2008 ) developed an intelligent real-time system for automatic weeding scheme in oil thenar plantation utilizing statistical attack GLCM and structural attack FFT and SIFT with a success rate above 80 % .

MATERIALS AND METHODS

The images to be used for this survey were taken from a chilli field. Besides five weed species were chosen which are common in chilli Fieldss of Bangladesh. Table I lists the English and Scientific names of chili and the selected weed species. Fig. 1 shows sample images of chili and other five weed species.

Table

SELECTED SPECIES

Class Label

English Name

Scientific Name

1

Chili

Capsicum frutescens

2

Lamb’s-quarters

Amaranthusviridis

3

Marsh herb

Enhydrafluctuans

4

Lamb ‘s quarters

Chenopodium album

5

Cogongrass

Imperatacylindrica

6

Burcucumber

Sicyosangulatus

( a ) ( B )

( degree Celsius ) ( vitamin D )

( vitamin E ) ( degree Fahrenheit )

Fig. 1 Sample images of different workss ; ( a ) chili ( B ) lamb’s-quarters ( degree Celsius ) fen herb ( vitamin D ) lamb ‘s one-fourth ( vitamin E ) cogongrass ( degree Fahrenheit ) burcucumber

Image Acquisition

The images were taken with an OLYMPUS FE4000 digital camera equipped with a 4.65 to 18.6 millimeter lens. The camera was pointed straight towards the land while taking the images. The lens of the camera was 40 centimeter above the land degree. An image taken with these scenes would cover a 30 centimeter by 30 centimeter land country. No flash was used while taking the image and the image scenes were protected against direct sunshine. The image declaration of the camera was set to 1200×768.The images taken were all RGB images.

Pre-processing

Segmentation method was used to divide the workss from dirt in images. Thresholding technique was used for this intent. The fact that workss are greener than dirt was used to make cleavage. Let ‘G’denotes the green coloring material constituent of a RGB image. A gray-scale image was obtained from the original image by sing merely the ‘G’value. A threshold value of ‘G ‘ was so calculated. Let ‘T ‘ denotes this threshold value. The pels with ‘G’value greater than ‘T ‘ were treated as works pels and lower than were soil pels. For each image, a binary image was obtained by cleavage, where pels with value ‘0 ‘ represent dirt and pels with value ‘1 ‘ represent works.

For taking noise from the images, an gap operation was foremost applied to the binary images. In gap, an eroding operation is followed by a dilation operation. It has the consequence of taking little pixel parts [ 5 ] . Then a shutting operation was applied. In shutting, a dilation operation is followed by an eroding operation. It will make full little holes in an object [ 5 ] .

( a ) ( B )

( degree Celsius ) ( vitamin D )

Fig. 2 Images of a lamb’s-quarters ; ( a ) RGB image ( B ) gray-scale image ( degree Celsius ) segmented binary image ( vitamin D ) binary image after noise remotion

Feature Extraction

A entire figure of 14 characteristics were extracted from each image. These characteristics can be divided into three classs: coloring material characteristics, size independent form characteristics and minute invariants.

Color Features: Let ‘R ‘ , ‘G ‘ and ‘B ‘ denote the ruddy, green and bluish colourcomponents severally. Every constituent was divided by the amount of all three constituents to do the coloring material features independent of different light conditions [ 6 ] .

R = ( 1 )

g = ( 2 )

B = ( 3 )

Merely works pels were used when ciphering the coloring material characteristics, so the characteristics are merely based on works coloring material non dirty coloring material. The coloring material characteristics used were: average value of ‘r ‘ , average value of ‘g ‘ , average value of ‘b ‘ , standard divergence of ‘r ‘ , standard divergence of ‘g ‘ and standard divergence of ‘b ‘ .

Size Independent Shape Features: The size independent characteristics used for this survey were:

Formfactor = ( 4 )

Elongatedness = ( 5 )

Convexity = ( 6 )

Solidity = ( 7 )

Here, country is the figure of pels with value ‘1 ‘ in a binary image.Perimeter is defined as the figure of pels with value ‘1 ‘ for which at least one of the eight neighbouring pels is a dirt pel. Thickness is twice the figure of shriveling stairss ( riddance of boundary line pels one bed per measure ) to do an object within an image disappear [ 7 ] . Convex country is the country of the smallest convex hull that contains all objects in an image. Convex margin is the margin of the convex hull that contains all objects in an image.

Moment Invariant Features: The minutes find how dispersed an object ‘s country is [ 6 ] . The undermentioned minute invariants were used:

I¦1 = I·2,0 + I·0,2 ( 8 )

I¦2 = ( I·2,0 + I·0,2 ) 2 + 4I·1,12 ( 9 )

I¦3 = ( I·3,0 – 3I·1,2 ) 2 + ( I·0,3 – 3I·2,1 ) 2 ( 10 )

I¦4 = ( I·3,0 + I·1,2 ) 2 + ( I·0,3 + I·2,1 ) 2 ( 11 )

Here,

( 12 )

where

( 13 )

and

( 14 )

degree Fahrenheit ( x, y ) is ‘1 ‘ for those braces of ( x, y ) that correspond to works pels and ‘0 ‘ for dirt pels. The minute characteristics are invariant to rotary motion and contemplation. The minute invariants were calculated on object country. Natural logarithm was used to do the minute invariants more additive.

Categorization Using Support Vector Machines

In SVM, a classii¬?cation undertaking normally involves dividing informations into preparation and proving sets. Each case in the preparation set contains one “ mark value ” ( i.e. the category labels ) and several “ properties ” ( i.e. the characteristics or ascertained variables ) . The end of SVM is to bring forth a theoretical account ( based on the preparation informations ) which predicts the mark values of the trial information given merely the trial information properties [ 8 ] . A preparation set of tuples and their associated category labels was used. Each tuple is represented by an n-dimensional characteristic vector,

wheren=14

Here, ‘X ‘ depicts n measurings made on the tuple from N features.There are six categories labelled 1 to 6 as listed in TABLE I.

SVM requires that each information case is represented as a vector of existent Numberss. As the characteristic value for the dataset can hold the value in dynamic scope, dataset demands to be normalized to avoid properties in greater numeral scopes ruling those in smaller numeral scopes. LIBSVM 2.91 was used for support vector categorization [ 9 ] . Each characteristic value of the dataset was scaled to the scope of [ 0, 1 ] . RBF ( Radial-Basis Function ) meat was used for SVM preparation and proving. As this kernel nonlinearly maps samples into a higher dimensional infinite so it can manage the instance when the relation between category labels and characteristics is nonlinear [ 8 ] . RBF meat requires two parametric quantities: ‘I? ‘ and a punishment parametric quantity, ‘C ‘ . Appropriate values of ‘C ‘ and ‘I? ‘ should be calculated to accomplish high success rate in categorization. For this survey, the values of these two parametric quantities were C = 1.00 and I?= 1 / entire figure of characteristics.

Result and Discussion

A common testing process is to divide the information set into two parts, of which 1 is considered unknown. The public presentation of sorting an independent testing informations set is reflected by the anticipation truth obtained from the “ unknown ” set. Cross proof is an improved version of this process which prevents the overi¬?tting job. Tenfold cross proof was selected for the proving intent. In tenfold cross-validation, the preparation set is divided into 10 subsets of equal size. Consecutive one subset is tested utilizing the classii¬?er trained on the staying nine subsets. Therefore, each case of the whole preparation set is predicted one time so the cross-validation truth is the per centum of informations which are right classii¬?ed. The cross proof consequence of the developed system utilizing all characteristics was 95.9 % over 224 samples.

All harvest images were identified right by SVM. But in instance of weed images, there were some misclassifications. Five images of lamb’s-quarters were misclassified as burcucumber. One image of burcucumber was misclassified as lamb’s-quarters. Two images of marsh herb were misclassified aslamb’squarters. No weed image was misclassified as Chilli. The overall categorization consequence is shown in TABLEII.

TABLE II

CLASSIFICATION RESULT USING ALL FEATURES

English Name of Samples

Number of Samples

Number of Misclassified Samples

Success Rate

Chili

40

0

100 %

Lamb’s-quarters

40

5

87.5 %

Marsh herb

31

2

93.5 %

Lamb ‘s quarters

33

0

100 %

Cogongrass

45

0

100 %

Burcucumber

35

2

94.3 %

Average Success Rate

95.9 %

To choose the set of characteristics which gives the best categorization consequence, both forward-selection and backward- riddance methods were used. In forward-selection, choice procedure starts with one characteristic and other characteristics are added one at a clip. At each measure, each characteristic that is non already in the set is tested for inclusion in the set. This procedure continues until no important betterment in categorization consequence is observed. In backward-elimination, the procedure starts with all characteristics included. At each phase, the least important characteristic is eliminated from the set. This procedure continues until a certain standard is met. These two procedures can be combined to happen an optimum set of characteristics. This procedure is called stepwise choice. In stepwise choice, characteristics are added as in forward choice, but after a characteristic is added, all the characteristics in the set are campaigners for backward-elimination. Using this method, a set of nine characteristics was selected which produce the best categorization rate. Those nine characteristics were:

Solidity

Elongatedness

Mean value of ‘r ‘

Mean value of ‘b ‘

Standard divergence of ‘r ‘

Standard divergence of ‘b ‘

log ( I¦1 ) of country

log ( I¦2 ) of country

log ( I¦4 ) of country

The consequence of tenfold cross proof utilizing these nine characteristics was 97.3 % . Four images of lamb’s-quarters were misclassified as burcucumber and two images of burcucumber were misclassified as lamb’s-quarters. All other images were classified accurately. The overall categorization consequence utilizing these nine characteristics is given in TABLEIII.

Table Three

CLASSIFICATION RESULT USING BEST FEATURES

English Name of Samples

Number of Samples

Number of Misclassified Samples

Success Rate

Chili

40

0

100 %

Lamb’s-quarters

40

4

90 %

Marsh herb

31

0

100 %

Lamb ‘s quarters

33

0

100 %

Cogongrass

45

0

100 %

Burcucumber

35

2

94.3 %

Average Success Rate

97.3 %

Decision

An automated weeding system must hold the ability to place harvests and weeds automatically and handle them consequently. Machine vision system based on digital image processing is found to be the most efficient detector sensing technique. For existent clip execution of this machine vision system, an efficient categorization theoretical account is required which can sort harvests and weeds with a high truth ratio. The end of this paper was to prove the feasibleness of support vector machines in harvests and weeds categorization. From the consequences, it is clear that SVM provides really high truth ratio and it is besides rather robust. To sort assorted weeds, farther research is required. One manner is to split the image into smaller parts and work with a individual part at a clip. Therefore, there will be less possibility to happen more than one weed classes in this little part.

Post Author: admin

x

Hi!
I'm Percy!

Would you like to get a custom essay? How about receiving a customized one?

Check it out