Scientia et Technica Año XXVIII, Vol. 29, No. 04, Octubre–diciembre de 2024. Universidad Tecnológica de Pereira. ISSN 0122-1701 y ISSN-e:2344-7214 158

Abstract— The present work presents the results of

research elaborated to recognize two classes of leaves of

weed present in the coffee crops through machine learning

techniques, a topic few have explored in the coffee

agroindustry, and that can significantly impact the

management of herbicides in this important crop. The study

involved twenty-four experiments, utilizing a database of

210 images, 70 for each weed class and 70 for coffee leaf

samples. All images were processed and transformed into

HSV color format. From each image, 33 texture patterns

were extracted and reduced to four through principal

component analysis. The fractal dimension was added as a

fifth pattern. The recognition used three machine learning

techniques: support vector machine (SVM), k-near

neighbors (KNN), and artificial neuronal networks. The

machine learning techniques permitted classification with

precision and recall upper or equal to 95%, on average,

when the fractal dimension was not used and upper or equal

to 97% on average when the fractal dimension was used.

SVM and ANN were methods with better outcomes. These

experiments constitute a first step towards implementing an

automatic system for selective weed eradication in a coffee

crop, with promising implications for future developments.

Index Terms—Coffee Crop, Weed Recognition, Texture

Analysis, Machine Learning.

Resumen— El presente trabajo presenta los resultados de

una investigación elaborada para reconocer dos clases de

hojas de maleza presentes en los cultivos de café mediante

técnicas de aprendizaje automático, un tema poco

explorado en la agroindustria cafetalera, y que puede

impactar significativamente el manejo de herbicidas en este

importante cultivo. El estudio involucró veinticuatro

experimentos, utilizando una base de datos de 210

imágenes, 70 para cada clase de maleza y 70 para muestras

de hojas de café. Todas las imágenes se procesaron y

transformaron en formato de color HSV. De cada imagen

se extrajeron 33 patrones de textura, que se redujeron a

cuatro mediante un análisis de componentes principales.

This manuscript was submitted on October 03, 2023. Accepted on November

13, 2024. And published on December 20, 2024.

La dimensión fractal se añadió como quinto patrón. Para el

reconocimiento se utilizaron tres técnicas de aprendizaje

automático: máquina de soporte vectorial (SVM), k-vecinos

más cercanos (KNN) y redes neuronales artificiales. Las

técnicas de aprendizaje automático permitieron la

clasificación con una precisión y exhaustividad superiores o

iguales al 95% en promedio, cuando no se utilizó la

dimensión fractal, y superiores o iguales al 97%, en

promedio, cuando se utilizó la dimensión fractal. SVM y

ANN fueron los métodos con mejores resultados. Estos

experimentos constituyen un primer paso hacia la

implementación de un sistema automático para la

erradicación selectiva de malezas en un cultivo de café, con

implicaciones prometedoras para desarrollos futuro.

Palabras claves— Cultivo de Café, Reconocimiento de Malezas,

Análisis de Textura, Aprendizaje Automático.

I. INTRODUCTION

EED control is crucial to ensure adequate growth and crop

performance. Indeed, weeds compete with crops for

water, nutrients, light, CO2, and space [1], with a special

affectation in the first plant years [2]. For coffee, one of the

most representative Colombian crops, this control must be

realized regularly throughout the year and imply considerable

spending of time and money. Anzalone and Silva ([3]) estimate

this spending to be around 35% when nothing is done and

between 16% and 27% when partial control is realized. The

conventional form of address weed eradication involves

applying chemically prepared herbicides for this purpose [4].

However, the regular or indiscriminate use of herbicides can

lead to severe environmental affectation, health problems, and

resistance of weeds to its application ([5], [6], [7], [8]). This

forces, in many cases, to alternate the use of glyphosate, which

is perhaps the most common herbicide used in coffee farms,

with other chemical herbicides ([9], [10], [11], [12], [13]).

Some possible secondary effects on the environment and

health are erosion, ground degradation, crop contamination,

water contamination, fauna, and human habitat contamination,

among others [14]. Roundup, for example, a widely used

herbicide in Colombia that contains glyphosate and

polyoxyethylene amine surfactant (POEA), is mentioned in

[15] as an herbicide capable of altering the hormonal balance of

the organisms that are exposed to it; its use is related to cases of

cancer and births with malformations. The affectation of flora

in coffee crops by herbicides is commentated by [16]. Damages

to earthworm biomasses, coleopterous, beneficial fungi, or

Weed Recognition in Coffee Crops Using

Texture Analysis and Machine Learning

Reconocimiento de Malezas en Cultivos de Café por Medio de Análisis de Textura

y Aprendizaje Automático

M. J. Muñoz-Neira

DOI: 10.22517/23447214.24589

Scientific and technological research paper

Scientia et Technica Año XXVIII, Vol. 29, No. 04, Octubre–diciembre de 2024. Universidad Tecnológica de Pereira.

159

edaphic mites are researched in [17], [18], [19], [20], [21], [22],

[23], [24]. Herbicides can even alter coffee crops themselves

[25], [26], [27].

The previous considerations show that the selective

application of herbicides or alternative methods for weed

eradication is fundamental. In both cases, the use of technology

for the automatic recognition of weeds is essential. Many works

have been developed for this purpose using different

techniques, like edge detection [28], image filtering [29],

hyperspectral sensing [30], crop signaling (making crops

distinguishable from weeds) [31], deep learning ([32], [33],

[34], [35], [36]), between others. Some research experiments

with systems for weed recognition in real-time [37], [38], too.

However, despite these works, there is no specific research

addressing the recognition of weeds in coffee crops.

The present paper summarized a series of experiments

tending toward recognizing images of two common weeds of

coffee crops, Sida Acuta and Paspalum Macrophyllum. Patterns

for recognition were obtained based on texture analysis, an

advantageous technique for leaf or fruit recognition ([39], [40]).

Three machine learning techniques were implemented: support

vector machine, SVM, k-near neighbors, KNN, and artificial

neuronal network, ANN, making use of Matlab 2022b®, with

the corresponding functions for these machine learning tools,

updated by the manufacturer for that version. Despite the

existence of more current and robust algorithms and

implementation tools, KNN, SVM, and ANNs remain highly

relevant and widely used in various applications ([41], [42],

[43], [44], [45]). Their enduring relevance is due to their status

as widely studied algorithms that are relatively easy to

understand and interpret. This makes them particularly suitable

for exploratory projects and with a reduced data set, such as the

one used in the present research.

The work constitutes a first approximation to the

implementation of an automatic system for selective weed

eradication. The paper is organized into four sections:

introduction, methodology, result analysis, and conclusion.

II. METHODOLOGY

Two hundred and ten samples of coffee plants and weed

leaves were recollected in an aleatory form from an area of 0.1

, in a coffee crop Castilla variety, in the farm San Carlos

from San Gil municipality, at the south of Santander, Colombia.

The samples were preserved by placing them into newspaper

sheets smeared with alcohol and transported to a laboratory

temperature-controlled at 20°C. Seventy leaves corresponded

to coffee plants, seventy to leaves of Sida Acuta weed, and

seventy to Paspalum Macrophyllum weed.

Each leaf was photographed with a digital camera in a light

cube, obtaining 210 RGB images of approximately 4608 by

3456 pixels, with 24 bits of profundity. This size was adequate

for image processing. Experiments with different image sizes

are beyond the scope set for investigation and may be the

subject of study in future work. The images were processed to

eliminate all that was not part of the leaf, resulting in images

like those shown in Fig. 1. Afterward, the processed RGB

images were transformed into HSI color format (hue,

saturation, intensity) through (1) to (3).

0, if Max

Min

―

Max

―

Min

4, if Max

―

Max

―

Min

2, if Max

―

Max

―

Min

if Max

(1)

(

Max

Min

)

(2)

0,if Max

Min (3)

Max

―

Min

Max

Min

Max

―

Min

, if Max

Fig. 1. Coffee and Weeds Leaves Images

Subsequently, from hue, saturation, and intensity

components, three co-occurrence matrixes (one for hue, the

other for saturation, and the other for intensity) were calculated,

according to (4). Eleven texture patterns were established from

each co-occurrence matrix according to (5) to (15).

Scientia et Technica Año XXVIII, Vol. 29, No. 04, Octubre–diciembre de 2024. Universidad Tecnológica de Pereira

160

∆𝑥∆𝑦

(

𝑖,𝑗

)

𝑛

𝑝

𝑛

𝑝

𝑖𝑓 𝐼

(

𝑝,𝑞

)

𝑖 𝑎𝑛𝑑 𝐼

(

𝑝

∆𝑥,𝑞

∆𝑦

)

𝑗

ℎ𝑒𝑟𝑤𝑖𝑠𝑒 (4)

Uniformity

𝐼

𝑁

𝑔

―

𝑖

𝑁

𝑔

―

𝑗

[𝑝(𝑖,𝑗)]

(5)

Meddle Intensity

𝐼

𝑁

𝑔

―

𝑖

𝑖𝑝

𝑥

(

𝑖

)

(6)

Moment Product

𝑁

𝑔

―

𝑖

𝑁

𝑔

―

𝑗

(

𝑖

―

𝐼

)

(

𝑗

―

𝐼

)

𝑝

(

𝑖

𝑗

)

(7)

Inverse Difference

𝐼

∑

𝑁

𝑔

―

𝑖

∑

𝑁

𝑔

―

𝑗

𝑝

(

𝑖,𝑗

)

(𝑖

―

𝑗)

(8)

Entropy

𝐼

𝑁

𝑔

―

𝑖

𝑁

𝑔

―

𝑗

𝑝(𝑖,𝑗

)

𝑝

(

𝑖,𝑗

)

(9)

Entropy Sum

𝐼

𝑁

𝑔

―

𝑖

𝑝

𝑥

𝑦

(𝑘)𝑙𝑛

𝑝

𝑥

𝑦

(

𝑘

)

(10)

Entropy Difference

𝐼

𝑁

𝑔

―

𝑖

𝑝

𝑥

―

𝑦

(𝑘)𝑙𝑛

𝑝

𝑥

―

𝑦

(

𝑘

)

(11)

Information Correlation 1

𝐼

―

𝐻𝑋𝑌

𝐻𝑋

(12)

Information Correlation 2

𝐼

―

𝑒

―

𝐻𝑋𝑌

―

𝐼

)

]

(13)

Contrast

𝐼

𝑁

𝑔

―

|𝑖

―

𝑗|

(𝑖

―

𝑗

)

𝑁

𝑔

―

𝑖

𝑁

𝑔

―

𝑗

𝑝(𝑖,𝑗)

(14)

Mode

𝐼

𝑁

𝑔

―

𝑖

𝑁

𝑔

―

𝑗

max[

𝑝(𝑖,𝑗)

]

(15)

Where p is the normalized co-occurrence matrix, p

is the

vector of the sum of columns of p, and p

is the vector of the

sum of rows of p. Values p

x+y

, p

x-y

, HX, HXY1 y HXY2, were

found conforming (16) to (20).

𝑝

𝑥

𝑦

(

𝑘

)

𝑁𝑔

―

𝑖

𝑁𝑔

―

𝑗

𝑝

(

𝑖,𝑗

)

𝑘

𝑖

𝑗

for k

0,1,2….2

𝑁

𝑔

―

(16)

𝑝

𝑥

―

𝑦

(

𝑘

)

𝑁𝑔

―

𝑖

𝑁𝑔

―

𝑗

𝑝

(

𝑖,𝑗

)

𝑘

𝑖

―

𝑗

for k

0,1,2….

𝑁

𝑔

―

(17)

𝐻𝑋

𝑁𝑔

―

𝑖

𝑝

𝑥

(

𝑖

)

∗

ln [

𝑝

𝑥

(

𝑖

)

]

(18)

𝐻𝑋𝑌1

𝑁𝑔

―

𝑖

𝑁𝑔

―

𝑗

𝑝

(

𝑖,𝑗

)

∗

ln [

𝑝

𝑥

(

𝑖

)

𝑝

𝑦

(

𝑖

)

]

(19)

𝐻𝑋𝑌2

𝑁𝑔

―

𝑖

𝑁𝑔

―

𝑗

𝑝

𝑥

(

𝑖

)

𝑝

𝑦

(

𝑖

)

∗

ln [

𝑝

𝑥

(

𝑖

)

𝑝

𝑦

(

𝑖

)

]

(20)

Next, a principal component analysis (PCA) was applied to

each group of patterns, including the group comprised of a mix

of all. The number of the first components selected (four) was

determined in a heuristic form. Consequently, for everyone, the

leaf was defined as four groups of patterns (H, S, I, and HSI),

with four patterns each.

Additionally, the fractal dimension of each image was

determined, subdividing the binarized image of each leaf in

boxes of increasing size, ri. The fractal dimension value was

obtained employing (21), where n is a vector in which each

element represents the number of boxes of dimension ri that

contain white color, and N is the length of the vector.

𝑑𝑓

∑

𝑖

∆

𝑛

𝑖

∆

𝑟

𝑖

𝑁

(21)

Three classifiers were used: quadratic support vector

machine (SVM), K nearest neighbors (KNN), and an artificial

neuronal network (ANN). SVM and KNN ran with a cross-

validation strategy and seven folds. SVM used a radial basis

Scientia et Technica Año XXVIII, Vol. 29, No. 04, Octubre–diciembre de 2024. Universidad Tecnológica de Pereira.

161

function as kernel, with box constraint and kernel scale of 1.

KNN used the five nearest neighbors classifier and the

Minkowski metric. ANN structure had two layers: the first with

neurons with sigmoid functions and the hidden layer with

softmax functions. Neurons in the hidden layer were five when

the number of inputs (patterns) was four and seven when the

number of inputs was five (with fractal dimension used as the

fifth pattern). All parameters of classifiers were adjusted in a

heuristic form.

ANN was trained with a scaled conjugate gradient

backpropagation algorithm. 70% of samples (147) were used

for training, 10% (21 samples) for validation, and 20% (42

samples) for testing.

With each classifier, eight essays were done, four without

fractal dimension (one for H patterns, another for S patterns,

another for I patterns, and another for HSI patterns), and four

adding the fractal dimension to each group of patterns. In total,

24 experiments were carried out. All codes and machine

learning tools were implemented in Matlab 2022b® in a

computer with an Intel Core i5 processor and 12GB of DDR4

memory.

III. RESULTS

For each experiment, the evaluation parameters described in

Table I were calculated. These are the same components of

matrix confusion but are presented in a table because, in this

form, it is easier to compare the results of the three systems

tested (SVM, KNN, and ANN). The total number of samples

determined the percentage of true or false positives. Thus, for

example, if the total true positives of coffee samples in SVM

classifications was 68, the percentage reported is 32.4% (that

corresponds to 68/210). Precision and Recall were estimated

according to (22) and (23), respectively, where TP is the total

of true positives, FP is the total of false positives, and FN is the

total of false negatives.

𝑃𝑟𝑒𝑐

𝑇𝑃

𝐹𝑃

(22)

𝑅𝑒𝑐

𝑇𝑃

𝐹𝑁

(23)

Precision indicates what percentage of the samples of the

systems identified as a positive class are true positives.

Precision is, therefore, a quality parameter. Recall indicates

what percentage of positive classes the systems could identify.

TABLE I

EVALUATION PARAMETERS

True Positives of Coffee

TPC

False Positives of Coffee by Weed One/False

Negatives of Weeds One by Coffee

FPC_W1

False Positives of Coffee by Weed Two/False

Negatives of Weed Two by Coffee

FPC_W2

True Positives of Weed One

TPW1

False Positives of Weed One by Coffee/False

Negative of Coffee by Weed One

FPW1_C

False Positives of Weed One by Weed Two/False

Negative of Weed Two by Weed One

FPW1_W2

True Positives of Weed Two

TPW2

False Positives of Weed Two by Coffee/False

Negatives of Coffee by Weed Two

FPW2_C

False Positives of Weed 2 by Weed One/False

Negatives of Weed One by Weed Two

FPW2_W1

Coffee Precision

Prec_C

Weed One Precision

Prec_W1

Weed Two Precision

Prec_W2

Coffee Recall

Rec_C

Weed One Recall

Rec_W1

Weed Two Recall

Rec_W2

The following tables (Table I to IX) synthesize the results

of the twenty-four experiments. The SVM column is the result

of the quadratic Support Machine Vector classifier, the KNN

column is the result of the fine K-Near Neighbors classifier, and

NNTr, NNV, and NNTe columns are the results of the training,

validating, and testing Neuronal Network classifier,

respectively.

TABLE II

EVALUATION PARAMETERS FOR HUE (H) TEXTURE PATTERNS

SVM

KNN

NNTr

NNV

NNTe

TPC

31,9%

31,0%

33,3%

19,0%

38,1%

FPC_W1

0,5%

1,4%

0,7%

0,0%

FPC_W2

1,0%

0,0%

TPW1

31,0%

31,4%

30,6%

42,9%

33,3%

FPW1_C

1,0%

0,5%

0,0%

4,8%

0,0%

FPW1_W2

1,4%

0,7%

0,0%

TPW2

32,4%

34,0%

33,3%

28,6%

FPW2_C

0,5%

0,0%

FPW2_W1

0,5%

0,7%

0,0%

Prec_C

95,7%

92,9%

98,0%

100,0%

Prec_W1

92,9%

94,3%

97,8%

90,0%

100,0%

Prec_W2

97,1%

98,0%

100,0%

Rec_C

95,7%

97,0%

100,0%

80,0%

100,0%

Rec_W1

97,0%

94,3%

95,7%

100,0%

Rec_W2

93,2%

98,0%

100,0%

For hue texture parameters, the three systems had results of

Precision and Recall up to 92%, being the best ANN.

TABLE III

EVALUATION PARAMETERS FOR SATURATION (S) TEXTURE PATTERNS

SVM

KNN

NNTr

NNV

NNTe

TPC

30,0%

25,7%

25,9%

14,3%

19,0%

FPC_W1

0,0%

0,5%

0,0%

2,4%

FPC_W2

3,3%

7,1%

23,1%

23,8%

31,0%

TPW1

32,4%

31,9%

32,7%

23,8%

35,7%

FPW1_C

0,5%

1,4%

0,7%

4,8%

0,0%

Scientia et Technica Año XXVIII, Vol. 29, No. 04, Octubre–diciembre de 2024. Universidad Tecnológica de Pereira

162

FPW1_W2

0,5%

0,0%

0,7%

0,0%

TPW2

32,9%

29,0%

9,5%

2,4%

FPW2_C

0,5%

4,3%

7,5%

19,0%

9,5%

FPW2_W1

0,0%

4,8%

0,0%

Prec_C

90,0%

77,1%

52,8%

37,5%

36,4%

Prec_W1

97,1%

95,7%

96,0%

83,3%

100,0%

Prec_W2

98,6%

87,1%

56,0%

28,6%

20,0%

Rec_C

96,9%

81,8%

76,0%

37,5%

66,7%

Rec_W1

100,0%

98,5%

100,0%

83,3%

93,8%

Rec_W2

89,6%

80,3%

28,6%

7,1%

In the case of saturation parameters, the performance was

lower than the results of hue parameters. For example, leaves

of coffee and weed number 2 were not recognized. The best

results were for the SVM system.

TABLE IV

EVALUATION PARAMETERS FOR INTENSITY (I) TEXTURE PATTERNS

SVM

KNN

NNTr

NNV

NNTe

TPC

32,4%

31,9%

32,0%

38,1%

31,0%

FPC_W1

0,0%

FPC_W2

1,0%

1,4%

0,0%

TPW1

32,4%

30,0%

22,4%

28,6%

26,2%

FPW1_C

0,0%

2,4%

FPW1_W2

1,0%

3,3%

4,8%

7,1%

TPW2

31,9%

28,6%

29,9%

23,8%

FPW2_C

0,5%

1,0%

0,7%

0,0%

FPW2_W1

1,0%

3,8%

10,2%

4,8%

9,5%

Prec_C

97,1%

95,7%

100,0%

Prec_W1

97,1%

90,0%

82,5%

85,7%

73,3%

Prec_W2

95,7%

85,7%

73,3%

83,3%

71,4%

Rec_C

98,6%

97,1%

97,9%

100,0%

92,9%

Rec_W1

97,1%

88,7%

68,8%

85,7%

73,3%

Rec_W2

94,4%

85,7%

86,3%

83,3%

76,9%

For intensity parameters, the results were better than those of

saturation parameters, globally. Better results were obtained

with the SVM method, with Precision and Recall up to 94%.

TABLE V

EVALUATION PARAMETERS FOR HSI TEXTURE PATTERNS

SVM

KNN

NNTr

NNV

NNTe

TPC

17,6%

13,6%

14,3%

19,0%

FPC_W1

10,0%

8,6%

15,0%

4,8%

19,0%

FPC_W2

5,7%

7,1%

8,2%

9,5%

TPW1

18,6%

21,9%

8,2%

0,0%

11,9%

FPW1_C

11,0%

6,2%

5,4%

4,8%

0,0%

FPW1_W2

3,8%

5,2%

4,8%

0,0%

2,4%

TPW2

11,0%

19,0%

21,1%

28,6%

16,7%

FPW2_C

15,7%

8,1%

15,0%

19,0%

9,5%

FPW2_W1

6,7%

6,2%

8,8%

19,0%

11,9%

Prec_C

52,9%

37,0%

50,0%

40,0%

Prec_W1

55,7%

65,7%

44,4%

0,0%

83,3%

Prec_W2

32,9%

57,1%

47,0%

42,9%

43,8%

Rec_C

39,8%

55,2%

40,0%

37,5%

66,7%

Rec_W1

52,7%

59,7%

25,5%

0,0%

27,8%

Rec_W2

53,5%

60,6%

62,0%

75,0%

58,3%

The machine learning process for the HSI texture parameters

was the worst performing, with Precision and Recall, on

average, lower than 50% for SVM, KNN, and ANN.

The following tables present the recognition results when a

fractal dimension is added. In general, as is evident from the

data, this addition improves the quality of machine learning.

TABLE VI

EVALUATION PARAMETERS FOR HUE (H) TEXTURE PATTERNS PLUS FRACTAL

DIMENSION

SVM

KNN

NNTr

NNV

NNTe

TPC

32,4%

31,4%

32,0%

33,3%

35,7%

FPC_W1

0,5%

1,4%

0,0%

4,8%

0,0%

FPC_W2

0,5%

0,0%

TPW1

32,4%

31,9%

33,3%

31,0%

FPW1_C

0,5%

1,0%

0,0%

2,4%

FPW1_W2

0,5%

0,0%

TPW2

32,4%

34,7%

28,6%

31,0%

FPW2_C

0,5%

0,0%

FPW2_W1

0,5%

0,0%

Prec_C

97,1%

94,3%

100,0%

87,5%

100,0%

Prec_W1

97,1%

95,7%

100,0%

92,9%

Prec_W2

97,1%

100,0%

Rec_C

97,1%

95,7%

100,0%

93,8%

Rec_W1

97,1%

94,4%

100,0%

87,5%

100,0%

Rec_W2

97,1%

100,0%

TABLE VII

EVALUATION PARAMETERS FOR SATURATION (S) TEXTURE PATTERNS PLUS

FRACTAL DIMENSION

SVM

KNN

NNTr

NNV

NNTe

TPC

32,4%

31,9%

28,6%

57,1%

35,7%

FPC_W1

1,0%

0,0%

FPC_W2

0,0%

0,5%

0,0%

TPW1

32,4%

31,9%

36,1%

28,6%

26,2%

FPW1_C

1,0%

1,4%

0,0%

2,4%

FPW1_W2

0,0%

TPW2

32,4%

35,4%

14,3%

0,0%

Scientia et Technica Año XXVIII, Vol. 29, No. 04, Octubre–diciembre de 2024. Universidad Tecnológica de Pereira.

163

FPW2_C

1,0%

0,0%

FPW2_W1

0,0%

35,7%

Prec_C

97,1%

95,7%

100,0%

Prec_W1

97,1%

95,7%

100,0%

91,7%

Prec_W2

97,1%

100,0%

Rec_C

94,4%

93,1%

100,0%

93,8%

Rec_W1

97,1%

100,0%

Rec_W2

100,0%

98,6%

100,0%

TABLE VIII

EVALUATION PARAMETERS FOR INTENSITY (I) TEXTURE PATTERNS PLUS

FRACTAL DIMENSION

SVM

KNN

NNTr

NNV

NNTe

TPC

32,4%

31,9%

31,3%

52,4%

28,6%

FPC_W1

0,0%

0,5%

2,7%

4,8%

0,0%

FPC_W2

1,0%

0,0%

TPW1

31,9%

31,4%

32,7%

19,0%

28,6%

FPW1_C

0,5%

0,7%

0,0%

FPW1_W2

1,0%

1,4%

0,0%

TPW2

32,4%

31,4%

32,7%

23,8%

40,5%

FPW2_C

0,5%

1,0%

0,0%

FPW2_W1

0,5%

1,0%

0,0%

2,4%

Prec_C

97,1%

95,7%

92,0%

91,7%

100,0%

Prec_W1

95,7%

94,3%

98,0%

100,0%

Prec_W2

97,1%

94,3%

100,0%

94,0%

Rec_C

97,1%

95,7%

97,9%

100,0%

Rec_W1

98,5%

95,7%

92,3%

80,0%

92,3%

Rec_W2

94,4%

93,0%

100,0%

TABLE IX

EVALUATION PARAMETERS FOR HSI TEXTURE PATTERNS PLUS FRACTAL

DIMENSION

SVM

KNN

NNTr

NNV

NNTe

TPC

25,7%

25,2%

29,3%

38,1%

21,4%

FPC_W1

7,1%

8,1%

8,2%

9,5%

2,4%

FPC_W2

0,5%

0,0%

TPW1

22,4%

23,3%

22,4%

14,3%

40,5%

FPW1_C

9,5%

8,6%

3,4%

0,0%

11,9%

FPW1_W2

1,4%

0,0%

TPW2

32,4%

31,0%

36,1%

38,1%

21,4%

FPW2_C

0,5%

1,0%

0,0%

FPW2_W1

0,5%

1,4%

0,7%

0,0%

2,4%

Prec_C

77,1%

75,7%

78,2%

80,0%

90,0%

Prec_W1

67,1%

70,0%

86,8%

100,0%

77,3%

Prec_W2

97,1%

92,9%

98,1%

100,0%

90,0%

Rec_C

72,0%

72,6%

89,6%

100,0%

64,3%

Rec_W1

74,6%

71,0%

71,7%

60,0%

89,5%

Rec_W2

94,4%

95,6%

100,0%

The results obtained are promising compared to those

presented by another related research on the subject. Although

there are no specific developments for weed recognition in

coffee crops in the literature consulted, it is possible to compare

the results obtained with those of works oriented, in general, to

weed recognition through different techniques. In [46], for

example, using convolutional neural networks, CNN, to

recognize different types of weeds, recognition percentages of

97.78% are achieved in validation. On the other hand, [47], an

exhaustive review of works on machine learning and deep

learning, highlights research results for crop and weed

discrimination with accuracy percentages of up to 95.1% using

ANN and up to 98.2% using SVM. In [48], for weed recognition

in strawberry and pea crops, recognition accuracies of 95.3%

are achieved with CNN and 63.7% and 84.9% for SVM and

KNN, respectively.

In synthesis, in this work, the experiments with only texture

patterns, hue (H), saturation (S), and intensity (I), without

considering fractal dimension, showed that quadratic SVM

reliably classifies coffee and weed samples, with Precision and

Recall near or upper to 95%, on average. KNN, for its part,

classifies with Precision and recall upper to 90% for Hue and

Intensity patterns and upper to 86% for S patterns, on average,

too; nevertheless, classification is better for SVM.

ANN does a reasonable classification for hue and intensity

patterns, but better for hue, with Precision and Recall equal to

100% in the testing samples of hue patterns and upper to 80%

for intensity patterns. Precision and Recall are minor to 60% for

saturation patterns, on average.

With Hue, Saturation, and Intensity (HIS) patterns, all

classifiers have Precision and recall minor to 60%, on average.

As was commented above, it is important to mention that

when the fractal dimension is added as the fifth pattern,

Precision and Recall improve in all classifiers, achieving values

near or upper to 97%, on average, in most

experiments. However, even if the fractal dimension is used,

the classifiers function better with individual patterns H, S, and

I than with pattern mixed (HIS). For pattern mixed (HIS) plus

fractal dimension, Precision and Recall achieve values of 80%,

on average.

IV. CONCLUSIONS

The extraction of texture patterns from coffee and weeds

images, in most experiments done, permits their classification

with precision and recall upper or equal to 95%, on average,

when the fractal dimension is not used, and upper or equal to

97% on average when the fractal dimension is used as the fifth

pattern. From the experimented classifiers, SVM and ANN

have better outcomes than the KNN method. The classification

has better results for individual texture patterns (hue, saturation,

intensity), reduced by PCA, than for the mixed pattern (HSI),

also reduced by PCA. In all tests, the fractal dimension

improves the performance of classifiers. Experiments suggest

that using this technology to identify and classify weeds

associated with the coffee crop is viable. Tests with an extended

group of weeds and samples, just as classification experiments

Scientia et Technica Año XXVIII, Vol. 29, No. 04, Octubre–diciembre de 2024. Universidad Tecnológica de Pereira

164

in real-time, are necessary for classifier

validation. Applications of this type of technology are

fundamental to improving the efficiency of the weed control

system for the benefit of coffee crops and environmental

protection.

REFERENCES

[1] Y. Blanco and Á. Leyva, “Las Arvenses En El

Agroecosistema Y Sus Beneficios Naturales,” Cultivos

Tropicales, vol. 28, no. 2, pp. 21–28, 2007, doi:

http://dx.doi.org/10.13140/RG.2.2.10964.19844.

[2] E. Gamboa and L. Sánchez, “Control de malezas con

herbicidad y métodos mecánicos en plantaciones

jóvenes de café,” Bioagro, vol. 16, no. 2, pp. 133–136,

2004.

[3] A. Anzalone and A. Silva, “Evaluación de herbicidas

sulfonilureas para el control de malezas en cafetales,”

Bioagro, vol. 22, no. 2, pp. 95–104, 2010.

[4] CENICAFE, “Manejo de las malezas o arvenses en los

cafetales,” 2004, CENICAFÉ.

[5] C. P. de Souza, T. de A. Guedes, and C. S. Fontanetti,

“Evaluation of herbicides action on plant bioindicators

by genetic biomarkers: a review,” Environmental

Monitoring and Assessment, vol. 188, no. 12, 2016, doi:

10.1007/s10661-016-5702-8.

[6] R. Hoagland and D. B. Clyde, “Controlling herbicide-

susceptible, -tolerant and -resistant weeds with

microbial bioherbicides,” Outlooks on Pest

Management, vol. 27, pp. 196–197, 2016, doi:

10.1564/v27_dec_04.

[7] W. T. Tsai, “Status of herbicide use, regulatory

management and case study of paraquat in Taiwan,”

Environment, Development and Sustainability, vol. 22,

no. 3, pp. 2673–2683, 2018, doi: 10.1007/s10668-018-

0293-x.

[8] M. Bilal, H. M. N. Iqbal, and D. Barceló, “Persistence

of pesticides-based contaminants in the environment

and their effective degradation using laccase-assisted

biocatalytic systems,” Science of the Total

Environment, vol. 695, 2019, doi:

10.1016/j.scitotenv.2019.133896.

[9] N. M. Freitas, F. C. L. Freitas, I. F. Furtado, M. F. F.

Teixeira, and V. F. Silva, “Herbicide mixtures to

control dayflowers and drift effect on coffee cultures,”

Planta Daninha, vol. 36, 2018, doi: 10.1590/S0100-

83582018360100047.

[10] P. V. da Silva, G. C. Barbosa, A. Ferrari, S. M.

Tronquini, and P. A. Monquero, “Chemical control

strategies of commelina benghalensis in coffee crop,”

Coffee Science, vol. 14, no. 2, pp. 231–239, 2019, doi:

10.25186/cs.v14i2.1576.

[11] X. M. de S. Vilela, F. C. L. de Medeiros, A. H.

Gonçalves, and L. C. da Silva, “Quizalofop-p-ethyl

controlling sourgrass (Digitariavilela insularis)x.

m.andde s. egoosegrass (eleusine indica) in infested

coffee areas,” Coffee Science, vol. 14, no. 4, pp. 530–

537, 2019, doi: 10.25186/cs.v14i4.1646.

[12] J. A. López, D. Villalba, L. F. Salazar, and O. A.

Cárdenas, “Manejo integrado de arvenses en el cultivo

de café Nueva alternativa de control químico,” Avances

Técnicos Cenicafé, pp. 1–8, 2012.

[13] A. Anzalone, M. Arizaleta, and J. Vargas, “Respuesta

del cafeto (Coffea arabica) ‘catuaí’ a los herbicidas

glifosato, clomazone, linuron, 2,4-D, metsulfuron-

metil, rimsulfuron y clorimuron-etil,” Bioagro, vol. 26,

no. 1, pp. 3–12, 2014.

[14] H. Menza and L. F. Salazar, “Estudios de resistencia al

glifosato en tres arvenses de la zona cafetera

colombiana y alternativas para su manejo,” Avances

Técnicos Cenicafé, no. 350, 2006.

[15] C. Martínez, “Taller sobre : “ Impactos toxicológicos y

ecotoxicológicos asociados ‘ Efectos tóxicos del

glifosato y sus formulaciones comerciales ,’” 2011,

Facultad de Agronomía Universidad de la República

de Uruguay.

[16] M. Arizaleta, A. Anzalone, and A. Silva, “Efecto del

uso de metsulfuron-metil y glifosato sobre malezas

asociadas a cafetales en venezuela,” Bioagro, vol. 20,

no. 2, pp. 79–88, 2008.

[17] J. A. García, E. Alarcón, Y. Hernández, and C.

Hernández, “Impact of litter contaminated with

glyphosate-based herbicide on the performance of

Pontoscolex corethrurus, soil phosphatase activities

and soil pH,” Applied Soil Ecology, vol. 104, no. 2015,

pp. 31–41, 2016, doi: 10.1016/j.apsoil.2016.03.007.

[18] R. Melloni, E. M. Silve, M. I. N. Alvarenga, E. G. P.

Melloni, and E. N. de Alcântara, “Impact of

weedcontrol on amf propagules and mycorrhization of

coffee trees,” Coffee Science, vol. 12, no. 2, pp. 207–

215, 2017.

[19] J. C. Niemeyer, F. B. de Santo, N. Guerra, A. M.

Ricardo Filho, and T. M. Pech, “Do recommended

doses of glyphosate-based herbicides affect soil

invertebrates? Field and laboratory screening tests to

risk assessment,” Chemosphere, vol. 198, pp. 154–160,

2018, doi: 10.1016/j.chemosphere.2018.01.127.

[20] P. De Pádua Marafeli, P. R. Reis, L. F. De Oliveira

Bernardi, E. N. De Alcântara, and P. A. Martinez,

“Effects of weed management on soil mites in coffee

plantations in a neotropical environment,” Neotropical

Biology and Conservation, vol. 14, no. 2, pp. 275–289,

2019, doi: 10.3897/neotropical.14.e38094.

[21] C. B. de Novais, L. Avio, M. Giovannetti, S. M. de

Faria, J. O. Siqueira, and C. Sbrana,

“Interconnectedness, length and viability of arbuscular

mycorrhizal mycelium as affected by selected

herbicides and fungicides,” Applied Soil Ecology, vol.

143, no. June, pp. 144–152, 2019, doi:

10.1016/j.apsoil.2019.06.013.

[22] S. Pochron et al., “Temperature and body mass drive

earthworm (Eisenia fetida) sensitivity to a popular

glyphosate-based herbicide,” Applied Soil Ecology,

vol. 139, no. March, pp. 32–39, 2019, doi:

10.1016/j.apsoil.2019.03.015.

[23] J. A. García-Pérez, E. Alarcón-Gutiérrez, and F. Díaz-

Fleischer, “Interactive effect of glyphosate-based

herbicides and organic soil layer thickness on growth

Scientia et Technica Año XXVIII, Vol. 29, No. 04, Octubre–diciembre de 2024. Universidad Tecnológica de Pereira.

165

and reproduction of the tropical earthworm

Pontoscolex corethrurus (Müller, 1857),” Applied Soil

Ecology, vol. 155, no. July 2019, p. 103648, 2020, doi:

10.1016/j.apsoil.2020.103648.

[24] Y. P. Setyawan, M. Naim, A. D. Advento, and J. P.

Caliman, “The effect of pesticide residue on mortality

and fecundity of Elaeidobius kamerunicus (Coleoptera:

Curculionidae),” IOP Conference Series: Earth and

Environmental Science, vol. 468, no. 1, 2020, doi:

10.1088/1755-1315/468/1/012020.

[25] G. Belutti et al., “Phytotoxicity and Growth of Coffee

Plants As a Function of The Application of Herbicide

2,4-D,” Coffee Science, vol. 14, no. 4, pp. 438–445,

2019.

[26] D. Toledo, A. de Oliveira, G. Belutti, T. Teruel, P.

Menicucci, and R. J. Guimarães, “Growth, anatomy

and physiology of coffee plants intoxicated by the

herbicide glyphosate,” Coffee Science, vol. 14, no. 1,

pp. 76–82, 2019, doi: 10.25186/cs.v14i1.1530.

[27] D. Toledo et al., “Selectivity of the herbicide

chlorimuron ethylcastanhon young coffee plants,”

Coffee Science, vol. 14, no. 4, pp. 467–472, 2019, doi:

10.25186/cs.v14i4.1615.

[28] L. Parra, J. Marin, S. Yousfi, G. Rincón, P. V. Mauri,

and J. Lloret, “Edge detection for weed recognition in

lawns,” Computers and Electronics in Agriculture, vol.

176, no. July, 2020, doi:

10.1016/j.compag.2020.105684.

[29] C. A. Pulido-Rojas, M. A. Molina-Villa, and L. E.

Solaque-Guzmán, “Machine vision system for weed

detection using image filtering in vegetables crops,”

Revista Facultad de Ingenieria, vol. 2016, no. 80, pp.

124–130, 2016, doi: 10.17533/udea.redin.n80a13.

[30] X. E. Pantazi, D. Moshou, and C. Bravo, “Active

learning system for weed species recognition based on

hyperspectral sensing,” Biosystems Engineering, vol.

146, pp. 193–202, 2016, doi:

10.1016/j.biosystemseng.2016.01.014.

[31] R. Raja et al., “Crop signalling: A novel crop

recognition technique for robotic weed control,”

Biosystems Engineering, vol. 187, pp. 278–291, 2019,

doi: 10.1016/j.biosystemseng.2019.09.011.

[32] A. dos Santos Ferreira, D. M. Freitas, G. G. da Silva,

H. Pistori, and M. T. Folhes, “Unsupervised deep

learning and semi-automatic data labeling in weed

discrimination,” Computers and Electronics in

Agriculture, vol. 165, no. July, 2019, doi:

10.1016/j.compag.2019.104963.

[33] T. Kounalakis, G. A. Triantafyllidis, and L.

Nalpantidis, “Deep learning-based visual recognition

of rumex for robotic precision farming,” Computers

and Electronics in Agriculture, vol. 165, no. August,

2019, doi: 10.1016/j.compag.2019.104973.

[34] J. Yu, S. M. Sharpe, A. W. Schumann, and N. S. Boyd,

“Deep learning for image-based weed detection in

turfgrass,” European Journal of Agronomy, vol. 104,

no. November 2018, pp. 78–84, 2019, doi:

10.1016/j.eja.2019.01.004.

[35] K. Hu, G. Coleman, S. Zeng, Z. Wang, and M. Walsh,

“Graph weeds net: A graph-based deep learning

method for weed recognition,” Computers and

Electronics in Agriculture, vol. 174, no. April, 2020,

doi: 10.1016/j.compag.2020.105520.

[36] H. Jiang, C. Zhang, Y. Qiao, Z. Zhang, W. Zhang, and

C. Song, “CNN feature based graph convolutional

network for weed and crop recognition in smart

farming,” Computers and Electronics in Agriculture,

vol. 174, no. April, p. 105450, 2020, doi:

10.1016/j.compag.2020.105450.

[37] R. Raja, T. T. Nguyen, D. C. Slaughter, and S. A.

Fennimore, “Real-time weed-crop classification and

localisation technique for robotic weed control in

lettuce,” Biosystems Engineering, vol. 192, pp. 257–

274, 2020, doi: 10.1016/j.biosystemseng.2020.02.002.

[38] R. Raja, T. T. Nguyen, V. L. Vuong, D. C. Slaughter,

and S. A. Fennimore, “RTD-SEPs: Real-time detection

of stem emerging points and classification of crop-

weed for robotic weed control in producing tomato,”

Biosystems Engineering, vol. 195, pp. 152–171, 2020,

doi: 10.1016/j.biosystemseng.2020.05.004.

[39] T. Ashraf and Y. N. Khan, “Weed density classification

in rice crop using computer vision,” Computers and

Electronics in Agriculture, vol. 175, no. June, 2020,

doi: 10.1016/j.compag.2020.105590.

[40] D. G. Kim, T. F. Burks, J. Qin, and D. M. Bulanon,

“Classification of grapefruit peel diseases using color

texture feature analysis,” International Journal of

Agricultural and Biological Engineering, vol. 2, no. 3,

pp. 41–50, 2009, doi: 10.3965/j.issn.1934-

6344.2009.03.041-050.

[41] A. Cravero, S. Pardo, S. Sepúlveda, and L. Muñoz,

“Challenges to Use Machine Learning in Agricultural

Big Data: A Systematic Literature Review,” Mar. 01,

2022, MDPI. doi: 10.3390/agronomy12030748.

[42] P. Dini and S. Saponara, “Analysis, design, and

comparison of machine-learning techniques for

networking intrusion detection,” Designs (Basel), vol.

5, no. 1, pp. 1–22, 2021, doi: 10.3390/designs5010009.

[43] C. Ji, T. B. Mudiyanselage, Y. Gao, and Y. Pan, “A

review of infant cry analysis and classification,” Dec.

01, 2021, Springer Science and Business Media

Deutschland GmbH. doi: 10.1186/s13636-021-00197-

[44] I. Attri, L. K. Awasthi, and T. P. Sharma, “Machine

learning in agriculture: a review of crop management

applications,” Multimed Tools Appl, vol. 83, no. 5, pp.

12875–12915, Feb. 2024, doi: 10.1007/s11042-023-

16105-2.

[45] M. J. Al Dujaili, A. Ebrahimi-Moghadam, and A.

Fatlawi, “Speech emotion recognition based on SVM

and KNN classifications fusion,” International Journal

of Electrical and Computer Engineering, vol. 11, no. 2,

pp. 1259–1264, Apr. 2021, doi:

10.11591/ijece.v11i2.pp1259-1264.

[46] Y. Mu et al., “DenseNet weed recognition model

combining local variance preprocessing and attention

mechanism,” Front Plant Sci, vol. 13, Jan. 2023, doi:

10.3389/fpls.2022.1041510.

[47] M. H. Saleem, J. Potgieter, and K. M. Arif,

“Automation in Agriculture by Machine and Deep

Scientia et Technica Año XXVIII, Vol. 29, No. 04, Octubre–diciembre de 2024. Universidad Tecnológica de Pereira

166

Learning Techniques: A Review of Recent

Developments,” Dec. 01, 2021, Springer. doi:

10.1007/s11119-021-09806-x.

[48] S. Khan, M. Tufail, M. T. Khan, Z. A. Khan, and S.

Anwar, “Deep learning-based identification system of

weeds and crops in strawberry and pea fields for a

precision agriculture sprayer,” Precis Agric, vol. 22, no.

6, pp. 1711–1727, Dec. 2021, doi: 10.1007/s11119-

021-09808-9.