Diferent diseases show diferent symptoms on plant leaves due to diferent reasons. Identifying plant leaf diseases without specialist expertise is typically challenging like finding region-areas of interest (ROIs) that is disease afected area on the leaf. The well known unsupervised machineilearningialgorithmsiK-meansiandiFuzzyiCMeansi(FCM) approaches based on clustering for plant leaf disease identification and severity evaluation is considered as a powerful technique to determine the form of illness has aflicted a particular tomato leaf and the severity. The segmentation technique is used to locate the ROI on a tomatoileaf. The objective of this clustering is to determine the accuracy of the segmented image of disease-afected leaves. The experiment utilizes images obtained from the Plant Village Dataset. The two stages of yellow leaf curl virus (YLCV) disease afected area have been observed. The severity score has been calculated for each type. The optimum cluster number of diferent severity level has been investigated based on various cluster validation index. Then diferent optimal cluster numbers of diferent severity levels have been shown by using Calinski and Harabasz index (CH) and Davies-Bouldin index (DB). The comparative study demonstrates the method's superiority relative to current methodologies.
A report published by the Food and Agriculture Organization[
In the 21st century, advanced technologies are being used in agriculture all over the world. In addition, to environmental and technical issues, some other issues like diverse soil and cropping patterns, digital divide i.e., unavailability of strong and stable internet connectivity, lack of confidence of farmers in new advanced technologies, lack of co-ordination between stakeholders which in turn highlights the issue of not reaching the farmers from diferent agricultural universities in the same geographical area at the same time and unreliable data sets respectively.
If the disease attacks the leaves of any tree, the colour and structure of the leaves of that tree will
change. Same things happens with tomato plants. For this reason, various researchers have studied
computer vision and machine learning algorithms for detecting leaf infections at diferent times of
diferent plants, classifying diseases and measuring disease severity. This section provides a various
computerivision and machineilearning based techniques used by researchers in plant disease detection
and disease severity measurement. Hong et al.[
Abisha & Jayasree[
Machine learning and deep learning have been recently emphasized in smart agriculture as tools for disease identification. Methods such as adaptive fuzzy systems and convolutional neural networks are used in previous works. However, unsupervised clustering for severity evaluation has received little attention in the literature. This study fills that need by using clustering algorithms and demanding evaluation metrics to validate the results. Innovative use of clustering indices and an emphasis on agriculturally-oriented, scalable systems are two important contributions.
The proposed methodology involves clustering-based image segmentation for disease severity evaluation. K-means and Fuzzy C-Means algorithms are used to identify afected regions. The methodology includes: 1. Data preprocessing to enhance image quality:
Images are preprocessed to enhance quality and ensure uniformity across the dataset. This includes resizing, normalization, and noise reduction.
2. Clustering to segment the diseased region. 3. Validation using Calinski-Harabasz and Davies-Bouldin indices to determine optimal clusters.
Comparative analysis with existing benchmarks demonstrates the approach’s efectiveness in identifying disease severity levels which is explained at the end of the section 5.
This work is proceeded up under unsupervised machine learning domain. The proposed method in this work evaluates the severity of tomato leaf disease without supervision, utilizing K-means and FCM clustering approach. Clustering is an unsupervised pattern classification method for machine learning. This clustering method is used as a very important tool in data science, where the input space can be divided into one or more clusters to identify the natural structure of data within a dataset[17]. Such an important clustering method is used in image processing, network sensing, pattern recognition, psychology, computer security, recommendation system, biology, text clustering etc. Hence, the detection of tomato leaf disease is also implemented by using clustering which is discussed in the following sections.
Let = {1, 2, . . . , } be a data set in a -dimensional Euclidean space R. Let = {1, 2, . . . , } be the cluster centers. Let = []× , where is a binary variable (i.e., ∈ {0, 1}) indicating if the data point belongs to the -th cluster ( = 1, . . . , ). The objective function of -means is as follows:
The -means algorithm is repeated through essential conditions for minimizing the -means objective function (, ) with updating equations for cluster centers and memberships, respectively, as: (, ) = ∑︁ ∑︁ ‖ − ‖2
=1 =1 = ∑︀=1 ∑︀ =1 (1) (2) = {︃1, if ‖ − ‖2 = min1≤ ≤ ‖ − ‖2
0, otherwise where ‖ − ‖ is the Euclidean distance between the cluster center and the data point .
A challenging issue in -means is that it requires specifying the number of clusters in advance, but this number is often unknown in practical situations. The superiority of the clusters produced by the -means algorithm depends on the initial cluster centers. This is also a limitation of using this algorithm. To resolve the issue of finding the number of clusters, cluster validity indices have gained much attention. Hence, Fuzzy C-Means (FCM) is used here for overlapping datasets.
FCM makes partition of a finite collection of elements = {1, . . . , } into a collection of fuzzy clusters based on certain principles. Given a finite set of data points from afected tomato leaves, the FCM algorithm produces a list of cluster centers = {1, . . . , } and a partition matrix: (3) (4) (5) (6) where is the hyper-parameter that controls how fuzzy the clustering will be. A higher value of results in fuzzier clusters.
The membership degree is calculated as: = 1
2 ∑︀=1 ︁( ‖− ‖ )︁ − 1
‖− ‖
After determining the number of clusters, cluster validation methodologies are applied. Two types of cluster justification indices, external and internal[ 18], are available. In this work, widely used internal indices such as Sum of Squares Between Clusters (SSB), Sum of Squares Within Clusters (SSW), CH[19], and DB[20] indices are employed to determine the number of optimal clusters and to analyze the relationship between optimal clusters and disease severity.
In the case of numerical data, the amount of SSW variance[21] can be determined with the help of the cluster centroid. SSW is used to measure cluster compactness. From the data obtained from various studies, it is observed that as the number of clusters increases, the value of SSW decreases. The equation to measure cluster compactness is given below:
= { | ∈ [
(, ) = ∑︁ ∑︁ ‖ − ‖2 =1 =1
SSW = ∑︁ ‖ − ‖2
=1
SSB is used to measure the degree of separation between clusters. The centroid distance from the mean vector of all objects is first measured, and this is used to compute the degree of separation between clusters. It is observed from the data of various studies[21] that as the number of clusters increases, the value of SSB also increases. The equation to measure the degree of separation between clusters is given below:
The value of CH is measured as the ratio of separation and compactness of the clusters. Cluster quality is better when the value of the CH index is maximum. SSB’s value should be higher, and the value of SSW should be lower as the number of clusters increases. The equation to measure the value of this index is given below:
SSB = ∑︁ ‖ − ¯‖2
=1 CH =
SSB/( − 1)
SSW/( − )
The DB criterion is based on a ratio of within-cluster and between-cluster distances. The equation to measure the value for this index is as follows:
SSW DB = (9)
SSB
Hence, by using these aforementioned approaches, the severity of disease on tomato leaves is identified, and the corresponding results and analysis are mentioned in the following section.
In order to measure tomato plant leaf disease severity, it is needed a large,iverifiedidataset ofiimages of healthy andidiseased tomato leaves. In this work, images of diseases of Tomato leaves have been taken from PlantVillage dataset[22]. There are 16,012 tomato leaves images in this dataset. The size of all the images is 256×256 and format is png. The PlantVillage data set contains a total of 1591 tomato healthy leaf images; out of 1591 only 50 images are randomly picked. The first six healthy leaf images from those are shown in Figure 1.
Similarly, 50 leaf images were randomly picked from 329 tomato yellow leaf curl virus (YLCV) leaf images in PantVillage dataset and stored in YLCV folder. Then from YLCV folder six images were picked in such a way that each leaf image has less yellow part than green part. These 10 leaf images are placed in the Grade-I folder which are shown in Figure 2.
The -Means clustering algorithm is applied on Sample-1 in Figure 1. A total of 9 cluster images with values of spanning from 2 to 10 are shown in Figure 4. The FCM clustering algorithm is applied on Sample-1 in Figure 1. A total of 9 cluster images with values of spanning from 2 to 10 are shown in Figure 5. The -Means clustering algorithm is applied on Sample-1 in Figure 2. A total of 9 cluster images are shown in Figure 6, with values of spanning from 2 to 10. The FCM clustering algorithm is applied on Sample-1 in Figure 2. A total of 9 cluster images with values of spanning from 2 to 10 are shown in Figure 7. Figure 3 shows the severity score of each sample image of Grade-I. Equation (10) is used to measure for each sample image belonging to Grade-I, where indicates the pixel number only within the regions of a disease-afected leaf, and indicates the total pixel number in that disease-afected leaf. (8) (10)
After obtaining the clustered image using K-Means and FCM clustering algorithms, internal validity index, DB is used to know how good the cluster structure is. Various graphs with optimum cluster number (k) obtained using DB are shown from Figures 12-17.
Clustering is commonly related to optimization problems by considering the optimization criteria to assess the class of clustering which belong to the internal validity index and related to static number of k clusters. Determining the optimal clusters number in a dataset is a very important task in partition clustering such as K-Means. SSW method is used to measure the compactness of a cluster. The smaller value of SSW indicates better quality in that cluster. Various experiments have shown that the value of SSW decreases as the clusters number increases. On the other side, SSB method is used to measure the separation between clusters. CH index is used to measure the best possible separation and compactness values between clusters. From Equation (8), it is easily understood that the worth of CH index increases when the value of SSB increases and the value of SSW decreases. So it can be said that with high separation the worth of CH index is maximum due to less error of compactness. The highest value of this CH index indicates the best clustering. On the other hand the value of DB index is shown in Equation (9). 50 Healthy leaves, 50 Grade-I diseased leaves, and 50 Grade-II diseased leaves (total 150 leaves) were used for generating CH index values as well as for DB index and then scatter diagrams are plotted in Figure 15 and Figure 16 respectively.
Figure 15 shows the scatter plot of the CH index. The CH index value was measured on images of 50 healthy leaves. A histogram of healthy leaves is plotted in Figure 17, where optimal k values along X-axis and the frequency of optimal values of k along Y-axis are shown respectively. Similarly, two more histogram plots are shown with Grade I and Grade II diseased leaves, in Figure 17. = 06 = 07 = 08 = 09 Figure 7: FCM Cluster images of diferent value of of − 1 of Figure 1. = 06 = 07 = 08 = 09 Figure 10: K-Means Cluster images of diferent value of of − 1 of Figure 3. − 1, * = 3 − 2, * = 3 − 3, * = 3 Figure 12: Samples of healthy tomato leaf images with DB index and optimum cluster number ( * ).
Below is a table presenting the comparative analysis between the proposed methodology and existing methods.
Key findings include: • Optimal clustering results with an efective segmentation of diseased regions.
− 1, * = 8 − 2, * = 8 − 3, * = 10 Figure 13: Samples of healthy tomato leaf images with CH index and optimum cluster number ( * ). − 1, * = 3 − 2, * = 2 − 3, * = 3 Figure 14: Samples of YLCV Grade-I tomato leaf images with DB index and optimum cluster number ( * ). − 1, * = 4 − 2, * = 4 − 3, * = 6 Figure 15: Samples of YLCV Grade-I tomato leaf images with CH index and optimum cluster number ( * ). − 1, * = 2 − 2, * = 3 − 3, * = 2 Figure 16: Samples of YLCV Grade-II tomato leaf images with DB index and optimum cluster number ( * ). • Comparative analysis highlights superior performance over existing methods.
− 1, * = 2 − 2, * = 3 − 3, * = 7
This work illustrates the usefulness of unsupervised machine learning methods, particularly K-means and Fuzzy C-Means (FCM), to determine disease severity in tomato leaves. The suggested method utilizes clustering-based image segmentation and validates the outcomes through Calinski-Harabasz and Davies-Bouldin indices, resulting in high accuracy in segmenting diseased areas. The comparative analysis demonstrates its superior efectiveness compared to existing approaches, especially in resource-constrained environments. These findings confirm the methodology’s potential for practical applications in smart agriculture. Subsequent study will concentrate on bringing together this methodology with Internet of Things (IoT) devices for continuous monitoring and decision-making in agriculture. Furthermore, overcoming limitations such as dependence on high-quality datasets and exploring its applicability to additional crops and diseases would improve its robustness and scalability. These innovations promise to enhance the accessibility and utility of precision agriculture, thereby largely contributing to sustainable farming techniques.
The author(s) have not employed any Generative AI tools. http://dx.doi.org/10.48084/etasr.3452. doi:10.48084/etasr.3452. [17] K. P. Sinaga, M.-S. Yang, Unsupervised k-means clustering algorithm, IEEE Access 8 (2020) 80716–80727. URL: http://dx.doi.org/10.1109/access.2020.2988796. doi:10.1109/access.2020. 2988796. [18] S. Lloyd, Internal versus external cluster validation indexes, IEEE Transactions on Information Theory 28 (2011) 129–137. URL: http://dx.doi.org/10.1109/tit.1982.1056489. doi:10.1109/tit. 1982.1056489. [19] T. Calinski, J. Harabasz, A dendrite method for cluster analysis, Communications in Statistics - Theory and Methods 3 (1974) 1–27. URL: http://dx.doi.org/10.1080/03610927408827101. doi:10. 1080/03610927408827101. [20] D. L. Davies, D. W. Bouldin, A cluster separation measure, IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-1 (1979) 224–227. URL: http://dx.doi.org/10.1109/tpami.1979. 4766909. doi:10.1109/tpami.1979.4766909. [21] Q. Zhao, Cluster validity in clustering methods, IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-1 (2012) 224–227. URL: https://doi.org/10.1109/tpami.2012.47909. doi:10.1109/ tpami.2012.47909. [22] Emmarex, Plant disease dataset, 2024. URL: https://www.kaggle.com/datasets/emmarex/ plantdisease, accessed on 1 September 2024.