Varietal Discrimination of Guava (Psidium Guajava) Leaves Using Multi Features Analysis

ABSTRACT

The purpose of this study was to examine the potential of Machine Vision (MV) approaches for the classification and identification of 12 varieties of guava. There are leaf images of the 12 local varieties of guava (Psidium guajava) that include Bangkok Red, China Surahi, Moti Surahi, Choti Surahi, Golden Gola, China Gola, Multani Sada Gola, Sadda Bahar Gola, Larkana Surahi, Black Guava, Hyderabadi Safeeda, Strawberry Pink Gola. A digital camera captured these images of guava varieties in a natural environment. Multi-features were extracted from these images. It was a composite of histograms, binary features, textures, rotational, spectral, and translational features (RST). Total 47 multi-features were collected for each non-overlapping guava leaf image, i.e., $\left(256\times 256\right)$ and $\left(512\times 512\right)$ more, the supervised correlation-based feature selection (CFS) method with the best search algorithm was used to optimize 18 features instead of 47 multi-features. Several ML classifiers, including Instant base Identifier (IBI), Random Forest (RF), and Meta Bagging (MB), using 10-fold cross-validation, were applied to the optimized multi-features. IBI results performed better than other classifiers with an average overall accuracy of 93.01% on AOIs,$\left(512\times 512\right)$. In addition, IBI detected 90.5%, 89.5%, 94%, 97%, 95.5%, 97%, 99%, 96.5%, 99%, 80.5%, 88%, and 81.5% accuracy values for the 12 varieties of guava leaves, namely Bangkok Red, China Surahi, Moti Surahi, Choti Surahi, Golden Gola, China Gola, Multani Sada Gola, Sadda Bahar Gola, Larkana Surahi, Black Guava, Hyderabadi Safeeda, Strawberry Pink Gola. The proposed study could play a significant role for the early and accurate identification of Guava varieties, and it would also be helpful for export quality measures for the national economy of the country.

KEYWORDS:

• Computer vision
• IBI
• Guava
• Pre-processing
• Classification
• Machine learning
• Medicinal plants

Introduction

Precision agriculture includes the accurate classification of plant varieties as one of its primary tasks. Even though plant discrimination requires human skills, knowledge, and expertise, this process is long, labor-intensive, and expensive. Each plant species has its own vein combination that distinguishes it from others and different texture combinations. It has been challenging for an expert’s eyes to distinguish different fruit varieties by observing plant leaves. A computer or mobile-held gadget’s eye (camera) equipped with artificial intelligence may quickly solve this complex problem with high accuracy without incurring huge costs. Several factors identify leaf phenotypes, such as color, shape, venation, tips, roughness, texture, size, and an edge map. The leaves serve as a fingerprint that identifies the type of guava plant.

Machine Vision (MV) techniques can reveal many aspects of natural science to laymen and enhance the expertise of professionals. Morphological features were used to identify plants using inherent methods. It has been found that visual^[1] and image analysis are practical tools for identifying diseases on strawberry leaves [anthracnose -resistant germplasm]. A quantitative and qualitative analysis of computer-based images was conducted to determine the percentage of damaged lesion areas on the infected leaves^[2] described a pattern recognition method for assessing disease severity and alerting farmers to save their crops.

The recognizing and detecting of the plants and their disease has been discussed in the literature using deep learning techniques.^[3] used 35000 infected and healthy leaves for detection and recognition. The mango varieties have been recognized through image processing and their disease identification^[4,5] The leaf damaged by different factors has been studied through machine learning algorithms. An analysis of multiclass datasets involving image processing methods and deep learning algorithms has been studied by^[6,7] An automatic leaf image-based cultivar label pipeline, namely MFCIS (Multi-Feature Combined Cultivator Identification System), collected through Permanent Homology and a Convolutional Neural Network (CNN) has been proposed by.^[8] The olive leaf classification based on their disease through profound convolutional neutral works has been described.^[9] Leaf stomatal density (SD) is measured by the number of stomata the leaf image has identified. Leaf thickness is also the feature of the leaf, which has been measured with the proposed model by.^[10] The application of deep learning agriculture has been reviewed by.^[11]

A variety of approaches, including fusion methods, are used to verify the identification of plants. To identify leaves and flowers, CNN models are used to combine leaf and flower images. The Kernal descriptor (KDES) was applied to leaf-based plant identification on scanned and natural images. KDES worked in three steps as Pixel level, Patch level, and Image level features for the extraction.^[12]

Plant leaf recognition through pattern recognition based on shape and multiple-texture features has been discussed in.^[13] Deep learning classification needs updating in its functionality and performance. ^[14,15] Plant phenotyping-driven studies have been well documented by^[16] Black rot and chlorosis disease of grapes leaves are identified through a transfer learning approach using a dataset of plant village images. Resnet50 framework is also applied to the same dataset. ^[17] Image classification of Papaya and disease on leaves are identified by the CNN ResNet50 method.^{[18 ,19]} In this study, they have used various factors to identify damaged areas of the left image, such as image size, image pose, illumination, complex background, different resolution, and image alignment. In addition to extracting guava seed oil, the composition of the fatty acids and their uses have been discussed in.^[16]

A guava fruit is considered the fourth (4th) fruit in the world. Guava is also known to be a major source of medical extracts. Among the benefits of guava leaves are anti-inflammatory, antibacterial, anti-oxidant, anti-cough, anti-diabetic, anti-diarrheal, anti-hyperlipidemic, cardio-tective, anti-mutagenic, larvicidal, and hepatoprotective effects. A guava is a fruit that grows in tropical and subtropical regions. Pink guava leaves possess many medicinal properties. The medicinal and nutritional properties of the guava have given the fruit pharmacological significance. It is helpful for the patient who has a deficiency of platelets in the blood and efficient result of recovery from dengue fever. Its fruit contains phytochemical elements, while the leaves contain bacterial and fungi static agents.^{[20, 21]} Guava leaves are rich in carotenoids, aldehydes, flavonoids, and tennis vitamins A and C. Its dried leaves powder is entirely medicinal for sugar patients.^[22] Pharankrang is also known as red guava. A morphological study of this plant and its future needs as a medicinal plant has been discussed.^[23] The principal yields attribute of guava leaf and its by-product through the principal component analysis has been discussed.^[24] Due to this, it has become an extraordinary fruit when compared to other types of fruits. Common guava, strawberry, lemon guava, pineapple guava, and apple guava with other local traits of the guava have been discussed. Pakistan, India, Sudan, Egypt, Indonesia, Brazil, Mexico, China, Philippines, Thailand, and Bangladesh are the guava-growing countries of the world. Guava is also considered a commercial plant from leaf to bark.^[25] Chemical leaf contents and remedies of the fruit drop of the guava plant have been discussed.^{[8, 11, 26–29]} Specifically, red guava leaves have a lot of phytochemical elements containing flavonoids, phenolic, tennis, and some essential oils explained in.^[30]

Pakistan is a leading exporter of guava pulp and fruit grown in the country. Generally, guava pulp comes in four colors: white, yellow, pink, and red. According to the Economic Wing of the Ministry of National Food Security & Research in Islamabad, Pakistan, April 2020. In Punjab province, Guava ranks sixth in terms of fruit production and has produced 447890 tons. Punjab’s According to the Economic Wing of the Ministry of National Food Security & Research in Islamabad, Pakistan, guavas ranks sixth in fruit production in Punjab province. There were approximately 46650 hectares of guava cultivation in Punjab in 2018. In the last five years, the yield has decreased due to fruit diseases, poor selection of plants, and low fruit quality and quantity. It is necessary to adopt technology (computer vision) to increase the yield of guava fruit. A major challenge facing computer vision and agriculturalists is classifying leaves, particularly guava leaves, which have a higher similarity percentage than other fruits. Domain and expert levels were characterized by complexity, time consumption, and inefficiency.

Guava leaves are largely similar and difficult to identify, even by experts. As detailed in the study, automatic systems are needed to identify guava leaves using fused methods. We can classify and discriminate guava leaves whose shape patterns are complex by using machine learning and deep learning frameworks. We propose a model that is not only useful to growers but also helpful to plant breeders, farmers, seed growers, and pulp manufacturers.

The present study contributes significantly to the field of guava and horticulture in Pakistan and around the world. The results of this study open the door to new research avenues in the field of guava leaves, stems, fruits, and related industries. Through our research farmers and pulp industry professionals will be able to grow better guava fruit by selecting the, best cultivars. This computer vision approach is intended to assist in developing mobile applications for detecting fruit leaves and early diseases. This timely action will assist growers, farmers, and the food industry for sustainable production of the guava fruit.

Materials and methods

This study illustrates the classification of different varieties of guava species with its local names in Pakistan are: Sada Bahar Gola, Moti Surahi, Chotti Surahi, China Gola, Sada Bahar Surahi, Golden Gola, Bangkok Red, Strawberry Pink Gola, Multani Desi Gola, Larkana, Black Guava, Hydrabadi Safeeda. All images have acquired from the fruit farms located at different cities of the Pakistan (Multan, Larkana, Sheikhupura, Lahore, Sharak pur shareef, Sargodha, Rahim Yar khan, Bahawalpur). These varieties are available in Pakistan on this website, https://aari.punjab.gov.pk/” The images were taken with a camera that has a resolution mega pixel. The color and fruit feature of local varieties available in Pakistan are mentioned in Table 1. Based on data collected from various guava farms, it has been observed that color is a weak distinguishing feature among varieties of guava. As a result of the analysis, it has been determined that the geometric and texture characteristics of the guava plant leaves are important characteristics for classification. The 200 fresh leaves image of each variety has been subjected to experimental testing after necessary pre-processing. We selected 12 varieties from a pool of locally available varieties.

Table 1. Local Guava Varieties Cultivar in Pakistan.

S. No	General Variety	Local Variety Fruit Features	Leaf Image	Images Taken
1	Red Bangkok Guava	Red skin and flesh with a sweet in taste. Rounded in shape.		1220
2	China Surahi	Crunchy and oval good shape. Full sweet in taste and more juicy. Yellowish skin with white yellowish pulp.		1340
3	Moti Surahi	Crunchy, oval, and bulk of irregular shape. Light sweet in taste, juicy. Green skin with white flesh.		1440
4	Choti Surahi	Small oval shape with sweet taste. Yellowish skin and white flesh. Bunch of fruit gives it beauty.		1305
5	Golden Gola	Green and yellowish skin and white flesh. Light sweet, and oval in shape.		1289
6	China Gola	Green and yellowish skin and white flesh. Light sweet, and rounded in shape. Heavy and loaded fruit in a bunch. Farmer like it due to its more and round-the-year production with heavy temperature resistance.		1378
7	Multani Sada Gola	Crunchy, rounded, and irregular shape. Sweet in taste, and less juicy with more seeds. Heavy weather resistance. Green skin with white flesh.		1390
8	Sadda Bahar Gola	Yellowish skin with white Flesh. Sweet in taste. More fruit loaded over on stem.		1089
9	Larkana Surahi	Oval in shape Sweet in taste and juicy. yellowish-white and crunchy flesh.		1270
10	Black Guava	Fruit is oval rounded and pear-shaped Medicated fruit. Black leaves fruit with black flesh with some seeds. Sour sweet with leaves and fruits are mildly fragrant,		1270
11	Hyderabadi Safeeda	Oval in shape Yellowish skin and white-yellowish flesh Sweet in taste.		1367
12	Strawberry Pink Gola	Strawberry shape with some pink spot. Green skin with pink spot and pink Flesh inside. Medicated fruit. Light sweet, lightly crunchy and juicy.		1280

Preprocessing of the images was carried out in the Anaconda environment using Python to resize images into various sizes like as $128\times 128,256\times 256,512\times 512,1024\times 1024$. Statistics and geometrical features are embedded with spectral features including histogram, second-order statistical texture, binary, rotational, scaling, and translational (RST) features. A fused feature data set of 47 features was developed for each sub-image. A total of $43200\left(18\times 2400\right)$ optimized feature spaces were developed through CFS implementation, and a multi-featured dataset for each size of AOIs was generated.

Proposed methodology

The proposed methodology is described in detail with steps. Image acquisition is followed by pre-processing carried out in the proposed system. A novel segmentation technique, Hybrid Threshold Range-Based Segmentation (HTRS), which is described in Algorithm 1, was successfully applied. Our proposed approach consists of three main steps: First, the HTRS technique was used to calculate the background pixels (BP) values at a particular threshold by considering a complete cluster as the image’s background. Second, this threshold level (BP) was used as a base pixel value to compare all adjacent pixels and complete the entire image. Finally, the gray level pixel value (GP) is considered a foreground pixel (FP) if the gray level value exceeds the background level (BP) and is considered the foreground or AOI. The steps of the proposed methodology are depicted in Figure 1.

Figure 1. Proposed Methodology of Guava Leaf Discrimination by Classification.

Multi-feature acquisition of the Guava leaf image

Multi-features were acquired for the classification of guava leaves in this study. A series of extracted features include: four first-order histograms, five higher-order texture features, 20 binary features, six spectral features, five geometric features, and seven RST features were extracted by each Area of Interest (AOI). A total of 47 features were extracted for each AOI. A total of 112800 (47x2400) feature space was developed.

First and higher order texture features of Guava leaf

Pixel Intensity value (PIV) is the primary information stored in a pixel of an image. These first and higher order statistical features can be derived by using the probability P(g) described in the equation (i(i) $P(g)=\frac{L(g)}{T}$(i) ) as:

(i) $P(g)=\frac{L(g)}{T}$(i)

Pixel intensity values were calculated by the five first-order histogram features that were Mean, Standard Deviation, Energy, Skewness, and Entropy. Mean describe the values of PIV in average form. In the above equation: T represents the Total Pixel in the image. L(g) represents the complete instance of the grayscale value of g.

(ii) $\bar{X}=\sum _{i=0}^{j-1}ij(i)=\sum _a\sum _b\frac{L(a,b)}{L}$(ii)

The J represents the gray-scale value range from 0 to 255 while a, and b represents the pixel in the area a and b on the leaf. The next equation is the standard deviation describing the image contrast as:

(iii) $SD(\sigma )=\sqrt{\sum _{i=0}^{k-1}(i-\stackrel{-}{i}}{)}^{2}k(i)$(iii)

When the center value is not found symmetric then the asymmetric skew was applied for the above images as in equation (iv)

(iv) $Skew=\frac{1}{\varphi \stackrel{3}{i}}\sum _{i=0}^{k-1}{(i-\bar{i})}^{3}k(i)$(iv)

The distribution of grayscale values is derived by the Energy as in equation (v(v) $Energy=\sum _{i=0}^{k-1}{\left[k(i)\right]}^2$(v) ).

(v) $Energy=\sum _{i=0}^{k-1}{\left[k(i)\right]}^2$(v)

The total images of the guava leaves describe the Entropy due to its disorder and uncertainty as in equation (vi)

(vi) $Entropy=-{\sum }_{j-0}^{r-1}r(j){log}_2\left[r(j)\right]$(vi)

The texture is a continuous repeating pattern in local variations of image intensity. During the experiment texture feature help out the partitioning of images into their region of interest and classify the specific regions for the identification of the varieties on the basis of arrangements of colors and intensities. Texture features are always based on some common factors such as the gray-level co-occurrence matrix (GLCM). For this study, five second = order features were studied in four different dimensions $(0^{\circ },{45}^{\circ },{90}^{\circ },{135}^{\circ })$ up to the −3-pixel difference, eight texture features were extracted: Energy, Entropy, Contrast, Homogeneity, Dissimilarity, Correlation, Inertia, and Inverse Difference. The following equations describe some experimental work as:

(vii) $P_d\left[i,j\right]=n_{ij}$(vii)

A GLCM, Pd[I,j] was defined as the first specifying displacement vector d = (dx, dy) and counted all the pairs of pixels separated by d having gray levels i and j.

$L_E=\sum _{i=1}^m\sum _{j=1}^n\left|C\right.\left.(i,j)\right|$(viii)

The gray level values obtained by the applying of the Energy. A gray level histogram of an image, measures the gray level. The x-axis shows a gray level from 0 to 255 while the y-axis shows the frequency of the pixels of the image. By using, Energy a threshold was calculated easily in the 12 varieties of guava leaves

(iX) $Entropy=-{\sum }_{j-0}^{r-1}r(j){log}_2\left[r(j)\right]$(iX)

Entropy measures the information contents of an image. Entropy was observed, highest when entries in x and y were in the same magnitude, and lowest when entries were found unequal.

(X) $Correlation=\frac{1}{{\varphi }_a{\varphi }_b}\sum _x\sum _y(X-{\eta }_a)(Y-{\eta }_b)Cxy$(X)

Correlation describes the pixel similarity for a particular pixel distance in the above equation (x).

$InverseDifference=\sum _m\sum _n\frac{C_{mn}}{\left|m-\left.n\right|\right.}$

The inverse difference was measured between m and n.

(xi) $C_{g=\sum _i\sum _j\hspace{0.17em}}\frac{P{\hspace{0.17em}}_d[i,j]}{1+\left|i\left.-j\right|\right.}$(xi)

A homogenous image result was in a co-occurrence matrix of guava (C_g) with a combination of high and low P[i,j]s.

(xii) $C(k,n)=\sum _i{\sum _j{(i-j)}^{k}P{\hspace{0.17em}}_d[i,j]}^n$(xii)

(xiii) $Inertia=\sum _i\sum _j{(i-j)}^2{C}_{ij}$(xiii)

By using of Contrast it was easy to measure the local variation in the image. Typically, the k = 2 and n = 1. This contrast describes the inertia.

(xvi) $SpectralRegionPower=\sum _{u\in \mathrm{R}\mathrm{e}gionv\in \mathrm{R}\mathrm{e}gion}\sum \left|T\left.{\left(u,v\right)}^2\right|\right.$(xvi)

Some spectral features were used during the classification of the guava leaf images based on texture in the form of power region. These power regions are called rings and the sectors as in the above equation (xvi), which helped in the above study.

Geometric features

The shape features are also called the binary features, such as axis of least second moments, area, Euler number, center area, aspect ratio, projection. Here 36 binary object features were extracted by projection of the normalization of height and width.

(xv) $Gi=\sum _{r=0}^{hight-1}\sum _{c=0}^{width-1}Ii(r,c)$(xv)

(xvi) $\bar{C}=\frac{1}{Gi}\sum _{r=0}^{hight-1}\sum _{c=0}^{width-1}cIi(r,c)$(xvi)

(xvii) $\bar{r}=\frac{1}{Gi}\sum _{r=0}^{hight-1}\sum _{c=0}^{width-1}rIi(r,c)$(xvii)

The centroid is the intersection of all hyperplanes of symmetry in the image as described in the above equation. The centers are defined by the row and column coordinates. The final orientation is defined in the angle of the axis relative to the vertical axis as described in the (xviii) equation as:

(xviii) $\theta i={Tan}^{-1}(2{\sum }_{r=0}^{height-1}{\sum }_{c=0}^{width-1}(r-\bar{r})(c-\bar{c})Ii(r,c)\left|{\sum }_{r=0}^{height-1}{\sum }_{c=0}^{width-1}(r-\bar{r}{)}^{2}Ii(r,c)\text{break}-\right.{\sum }_{r=0}^{height-1}{\sum }_{c=0}^{width-1}(c-\bar{c}{)}^{2}Ii(r,c)$(xviii)

Boundary of the Guava leaf image

Boundaries of the guava leaf images were derived by using the perimeter that defines the total pixels that make the boundary of the image. It was helpful to find the shape also. By labeling it, it counts the number of the pixel ‘1’ with neighboring pixel ‘0.’ The edge detector was used to measure pixel ‘1.’ The thinness ratio was calculated by the ratio of the area and perimeter as described in the equation (xix).

(xix) $Ti=(Gi{\left|Pi\right.}^2)$(xix)

The elongation was observed by the aspect ratio by finding the minimum and maximum values in the columns and rows as shown in the equation (xx). The aspect ratio is the proportional relationship between the width and height of the image.

(xx) $AspectRatio=\frac{C_{max}-C_{min}+1}{G_{max}-G_{min}-1}$(xx)

Topology of the Guava leaf image

The Euler number is the measure of the topology of the image. It describes the total number of the objects and number of the holes in the image.

Euler Number = (Total Number of the Objects – Total Number of the Holes) (xxi)

Projection of the leaf image

The projection describes the shape of the image. The projection profile of an image which one in particular direction always refers to the running sum of pixels in the same direction. This feature helps out to find the varieties of the same species which has a very minor difference. The projection profile basically converts the grayscale image to a binarized/threshold image. The vertical and horizontal projection is shown in the equations (xxii) and (xxiii).

(xxii) $hi(r)=\sum _{c=0}^{width-1}cIi(r,c)$(xxii)

(xxiii) $Vi(r)=\sum _{r=0}^{height-1}cIi(r,c)$(xxiii)

Feature deduction

It was observed that all the extracted features were not equally important for each variety of leaf image. During experimentation, it was noticed that to handle a large dataset was not so easy task for feature optimization purpose and to reduce the dimensionality of large feature space, we have deployed co-relation-based feature selection (CFS) with the best search algorithm on original guava leaf feature space (112800). A reduced 18 multi-feature dataset was acquired, instead of 47 features dataset for each image. A total number of 43200 (18 × 200) optimized feature space was developed by implementing CFS with the best search algorithm. During the optimization process, we have produced the feature space dimensionality without changing the original nature of the guava image’s feature space Table 3. Table 2 describes the list of optimized features

Table 2. Hybrid Threshold Range-Based Segmentation (HTRS).

Algorithm 1: Hybrid Threshold Range Based Segmentation (HTRS) Input: Guava Image Dataset Parameter: GID {Guava Image Dataset} Output: Cluster with Area of Interest (AOI)
1:	Function HTRS(GIDataset)
2:	Begin
3:	For each image in GIDataset do
4:	GTR←compute the total pixel range
5:	GBP←calculate the guava image base hybrid range
6:	If GTR≠GBP, then
7:	GPC←find guava pixel cluster
8:	AOI←find area of interest
9:	end if
10:	end for
11:	return {GPC,AOI}
12:	end
13:	end function

Table 3. Correlation-based Feature Selection (CFS) List.

S.No	Features
1	Object-Column Coordinate
2	Area
3	Centroid (Row_Column)
4	Orientation (Axis_of_Least_Second_Moment
5	Histo-Mean
6	Histo-Standard Deviation
7	Histo_Skew
8	Histo_Energy
9	Histo_Entropy
10	Texure_enrgy_average (Texure_Distance:2_IndudeZeroPairs:False)
11	Texure_enrgy_range
12	Inertia_average (Texture_Distance:2_IndudeZeroPairs:False)
13	Correlation_range
14	Inverse_diff_average (Texture_Distance:2_IndudeZeroPairs:False)
15	Invers_diff_range
16	Texture_entropy_average (Texure_Distance:2_IndudeZeroPairs:False
17	Texture_entropy_range
18	Spectral_DC

Results and discussions

In this study, we applied three machine vision classifiers, namely, Instant base Identifier (IBI), Random Forest (RF), and Meta Bagging (MB) to conduct a comparative analysis of twelve (12), Guava (Psidium Guajava) plant varieties using a fused multi-feature dataset. The proposed result was compared with the review results mentioned above. It was observed that the fusion approach gives outclass results for identifying guava varieties. The leaf images of guava varieties are more similar in their shapes and texture. They cannot be easily identified by the naked eye and simple classifier in their small sizes. For the Area of interest (AOI) the image’s pixel sizes were $64\times 64and128\times 128$, the classifiers failed to provide better accuracy and required results were less than 20% accuracy. To acquire better accuracy results, it was decided to increase the AOI size, then we used $256\times 256$ and applied the same three classifiers as discussed above in this section. After deploying these classifiers on 256 × 256, we had observed that improvement in results was expectedly were seen. The deployed classifiers, namely IBI, RF, and Meta Bagging were produced accuracy 87.125%, 85.625%, and 80.00%, respectively. These results were satisfactory as per previously acquired results but improvement can be brought by increasing the size of 512 × 512 on same classifiers. It was observed that the results were out performed on AOI size of 512 × 512 and achieved an accuracy of 93.00%, 90.03%, 82.75%, IBI, RF, Meta Bagging, respectively. Furthermore, we tried to enhance the accuracy on the AOI size of (1024 × 1024) but the results dramatically decreased to almost less than 20%. During these experimentations, it was noticed that as AOI size increase and decrease up to some particular extent, It results also variate or change exponentially. Now a specific size of AOI produced better results for the classification. We did perform five experiments on the sample data with a 10-fold cross-validation method. The training and testing data distribution was shown in detail in Table 4 while image size of (256 × 256) results is in given below Table 5.

Table 4. Training and Testing Data Distribution for Experiments.

Number of Varieties	Images Per Variety	Total Images	Per Image Derived Features	Total Derived Features in Images	Per Image Optimized Features	Total Optimized Features
12	200	2400	47	112800	18	43200

Table 5. Machine Learning Classifier Performance on Guava Leaves Images of Size (256 × 256).

Classifier	KappaValue	True Positive	False Positive	Receiver Operating Characteristics ROC	Mean Absolute Error MAE	Root- MeanSquare Error	Time (Sec)	Accuracy (%)
IBI	0.8595	0.871	0.012	0.93	0.0215	0.1465	0.01	87.125
RF	0.8432	0.856	0.013	0.989	0.0532	0.1439	3.76	85.625
Meta Bagging	0.7814	0.018	0.8	0.981	0.0653	0.1638	1.09	80.00

The different performance measuring parameters like Kappa value, True Positive, False Positive, Receiver-operating characteristics, Mean Absolute Error Values, Root Mean Square Error, Time, Accuracy of image size 256 × 256 illustrated in Figure 2. The confusion matrix is shown in Table 4. The confusion matrix graph of IBI of size (256x256) is shown in Figure 3.

Figure 2. Machine Learning Classifier Performance on Guava Leaves Images of Size (256 × 256).

Figure 3. Confusion Matrix Graph of IBI Classifier on Leaves Image Size (256 × 256).

The machine learning classifier performance on Guava leaf images of size (256 × 256) are shown in Table 6.

Table 6. Confusion Matrix Table of IBI Classifier on Leaves Image Size (256 × 256).

Classes	Bangkok Red	China Surahi	Moti Surahi	Choti Surahi	Golden Gola	China Gola	Multani Sada Gola	Sada Bahar Gola	Larkana Surahi	Black Guava	Hyderabadi Safeeda	Strawberry Pink Gola
Bangkok Red	154	34	11	1	0	0	0	0	0	0	0	0
China Surahi	40	126	26	1	0	2	0	1	3	1	0	0
Moti Surahi	12	25	147	0	0	1	0	0	0	2	2	11
Choti Surahi	0	0	0	195	0	3	1	0	1	0	0	0
Golden Gola	0	0	0	2	192	5	0	1	0	0	0	0
China Gola	0	0	0	0	6	194	0	0	0	0	0	0
Multani Sada Gola	0	0	0	0	1	0	198	0	0	0	1	0
Sadda Bahar Gola,	0	0	1	0	1	0	0	197	0	0	1	0
Larkana Surahi	0	0	0	0	0	0	0	1	198	0	0	1
Black Guava,	0	0	0	0	0	0	0	2	0	162	12	24
Hyderabadi Safeeda	0	0	1	1	0	0	1	1	1	7	176	12
Strawberry Pink Gola	0	0	10	1	0	0	0	0	0	25	12	152

The machine learning classifier performance on Guava leaf images of size (512 × 512) are shown in Table 7. The different performance measuring parameters like Kappa value, True Positive, False Positive, Receiver operating characteristics, Mean Absolute Error Values, Root Mean Square Error, Time, Accuracy of size 512 × 512 are shown in Figure 4: The Machine learning confusion matrix on Guava leaves images of size (512 × 512) are shown in Table 6 while graphically performance is shown in Figure 5.

Table 7. Machine Learning Classifier Performance on Guava Leaves Images of Size (512 × 512).

Classifier	Kappavalue	True positive	False positive	Receiver operating characteristics roc	Mean absolute error mae	Root- meansquare error	Time (sec)	Accuracy (%)
IBI	0.9164	0.923	0.007	0.958	0.0128	0.113	0.15	93.00
RF	0.8909	0.9	0.009	0.992	0.0483	0.1293	0.1004	90.03
Meta Bagging	0.8118	0.828	0.016	0.981	0.0669	0.1614	0.75	82.75

Figure 4. Confusion Matrix Graph of IBI Classifier on Leaves Image Size (512 × 512).

Figure 5. Confusion Matrix Graph of IBI Classifier on Image Size (512 × 512).

The Comparison of proposed approaches with existing approaches is shown in Table 9 for guava leaf dataset. The confusion matrix shows the result of 12 varieties of guava plant, namely, Bangkok Red, China Surahi, Moti Surahi, Choti Surahi, Golden Gola, China Gola, Multani Sada Gola, Sadda Bahar Gola, Larkana Surahi, Black Guava, Hyderabadi Safeeda, Strawberry Pink Gola were 90.5%,89.5%,94%,97%,95.5%,97%,99%,96.5%,99%,80.5%,88% and 81.5% in Table 8. The Diagonal Values are shaded in the confusion matrix diagonally. Our results were found with better results as compared to the classifiers used before on the different datasets of Flavia, Swedish, LifeClef, and Plant Village datasets. The success of our proposed model is attributed to a few fundamental and prominent factors.

• Pre-processing and enhancement of the acquired data to improve accuracy

• The multi-feature dataset was developed using fusion methods of the computer vision framework.

• Leaf Shape, Texture, Blade Area, and Venation parameters helped to achieve the desired results.

• The venation and texture of the leaves were depicted using images of 300–600 dpi.

• The background of the leaves was removed in the pre-processing stage, which helped with Blade area findings of 12 species.

Table 8. Confusion Matrix Table for IBI classifier on Image Size (512 × 512).

Classes	Bangkok Red	China Surahi	Moti Surahi	Choti Surahi	Golden Gola	China Gola	Multani Sada Gola	Sadda Bahar Gola,	LarkanaSurahi	Black Guava,	Hyderabadi Safeeda	Strawberry Pink Gola
Bangkok Red	181	5	8	0	0	0	0	2	1	1	1	1
China Surahi	7	179	10	1	0	0	0	1	0	2	0	0
Moti Surahi	3	5	188	0	1	0	1	0	0	0	2	0
Choti Surahi	0	1	1	194	0	2	1	0	1	0	0	0
Golden Gola	0	0	0	2	191	5	1	0	0	0	1	0
China Gola	0	0	0	0	6	194	0	0	0	0	0	0
Multani Sada Gola	1	0	0	0	1	0	198	0	0	0	0	0
Sadda Bahar Gola,	0	1	0	0	1	0	0	193	1	0	2	2
Larkana Surahi	0	0	0	0	0	0	0	1	198	0	1	0
Black Guava,	0	0	1	0	0	0	0	2	0	161	10	26
Hyderabadi Safeeda	1	0	1	1	0	0	0	1	1	7	176	12
Strawberry Pink Gola	0	0	0	0	0	0	0	0	0	24	13	163

Table 9. Comparison of Proposed Approaches with Existing Approaches Working on Leaf Dataset.

Authers Year	Dataset	Categories	Features	Classifiers	Accuracy [%)
.13	Swedish Leaf	15	Combined Feature [Shape+Texure]	MTD+LBP+HF	85.74
8	Sweet Cherry Leaf Dataset	88	Shape + Texture+Venation	DCNN+Xception/MFSCIS	83.52
,31	Swedish Leaf	12	Texture	K-NN,Fuzzy	85.75
32	Ocotea [RNV]	31	Venation Pattern	Faxitron DX,	86.00
] 33	Self-Acquired Dataset	7	Color+Shape	ANN+MLP+Neuro-Fuzzy	88.43
34	Self-Acquired Dataset	4	LBP,LGP,NABP	DCNN,SVM	98
35	Self-Acquired Dataset	14	Leaf Phenotype	CRD	67
36	Self-Acquired Dataset	40	Color+Shape, Leaf Tip,Leaf Margin	PC,PCA,	79
37	Self-Acquired Dataset	5	Color + Shape	CNN	70
38	LifeCLEF-2015	7	Shape+color+Texture	ImageCLEF,KNN	97
39	PlantVillage,	4	Shape +Texture	KNN,CNN	95.
Proposed approach	Self-Acquired Dataset of Guava Leaves	12	Shape +Texture	IBI, RF, Meta Bagging	93.00

The following graph in Figure 6 demonstrates the best comparison of the classifier IBI in different sizes of the images include AOI (256 × 256) and (512 × 512).

Figure 6. The Comparison Analysis Graph, on Twelve Varieties of Guava Leaves Images between (256 × 256) and (512 × 512).

Plants play an essential role in medicine, food, industry, and the protection of the environment. Recognizing plants, is beneficial in the agricultural industry from the onset of planting to harvesting the final product. However, identifying plant species that are similar to one another is challenging since it requires a special skill or a more excellent knowledge of that particular plant species. A similar objective to that of the above study was to be achieved. There is a need and value for an automated classification and identification system for plant species since the system can assist specialists and the public in identifying plant species as quickly as possible.

Conclusion

This study focused on classifying “12” different guava fruit plant leaves that describe their diversities, using the most effective Machine Learning (ML) algorithms. The best classifiers were used to accomplish this, including IBI, RF, and Meta Bagging. It has been illustrated that the results of all the classifiers are different. However, IBI performs well, achieving 93.00% accuracy. As a result, the proposed method seems suitable for real-time applications in agricultural fields. The study also led to the development of a mobile application to identify guava leaves for farmers for better production and selection of varieties. There may be an opportunity to utilize data expansion techniques in the future to improve classification accuracy.

Disclosure statement

No potential conflict of interest was reported by the authors.

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.