Many researchers have previously worked to automatically and accurately diagnose illnesses using various classification approaches, We have surveyed such related papers to our work here. Youwen et al. (2008) detected downy mildew and powdery mildew diseases on cucumber plants. They segmented the leaf photos using statistical patterns and mathematical morphology after using a median filter to reduce the noise. They used SVM classification for disease detection. The downy mildew and powdery mildew diseases were detected using the features taken from the diseased cucumber leaf image and fed to the SVM classifier [7].
Arivazhagan et al. (2013) worked on identifying diseases such as late scorch, Leaf Lesion, bacterial spot, Ashen Mold, Fungal spot, Scorch, sunburn, Late blight, early Blight, and sooty mold, early scorch, brown spots, yellow spots, bacterial diseases, late scorch, and fungal diseases on plant leaves such as Banana, tomato, Beans, Potato, Mango, Lemon, and Jackfruit. A digital camera takes pictures of different leaves to determine the illnesses from the indications of the leaves. By masking and removing the green pixels from digital RGB photos of leaves, HSI color images were created. [22] using a specified threshold value, segmenting the image, and computing the texture surface statistics. They utilized an SVM which made use of textural information to categorize the illness. They claimed that while identifying disease infection in plant leaves, an accuracy of 94.74% was attainable [8].
H. Al-Hiary et al. (2011) employed the Otsu segmentation method and K-means clustering approach to locate the areas of plant and stem illnesses that infect a different plant leaf. For the sake of extracting the characteristics of diseased sections, they employed the color co-occurrence approach. They identified diseases such as Early Scorch, White Mold, Tiny Whiteness, Cottony Mold, Late Scorch, and Ashen Mold using an ANN classifier [9,24].
Sannakki et al. (2013) took pictures of Grape plant leaves and identified the illness using image processing and artificial intelligence techniques. They categorized grape leaf diseases as Downy and Powdery Mildew. To increase the accuracy. Masking was used to get rid of the backdrop.
Anisotropic diffusion was used to preserve the information from the damaged leaf section. The authors applied K-Means as a segmentation technique, and GLCM was used for feature extraction. In the end, a classifier known as the Feed Forward Back Propagation Network classifier performs classification. [10].
Hossain et al. (2019) used the KNN classification approach to identify the diseases such as Alternaria Alternate, Leaf Spot, Anthracnose, Canker, and Bacterial Blight. For the input, 237 leaf photos were obtained from the database of Arkansas plant disease. The segmentation, followed by the GLCM matrix, was used to extract features from photos of sick plants, and the training dataset was subjected to 5-fold cross-validation to avoid overfitting. This process offered 96.76% accuracy [11].
Samajpati et al. (2016) employed the Random Forest classifier to recognize illnesses of the apple fruit, such as Rot, Scab, and Blotch. K mean clustering was used for image segmentation to find contaminated areas, and the fusion approach was used to extract features. Feature level fusion increased the illness classification's accuracy [12].
Pang et al. (2011) endeavoured to detect the illnesses like Cercospora Leaf Spots, Leaf Sheaths, and Brown Spots that have been detected in maize crops. The method found that all pixels for which the green channel's (G) and red channel's (R) and grey levels were more significant than each other. The diseased region consists of 98% red pixels; corresponding regions were detected, appropriately named, and used as seeds in a region-growing approach to more reliably and effectively detect diseased regions to the level of 96% [13].
Ramesh et al. (2020) focus on GLCM feature extraction and k-means clustering for precise position detection in plant leaves. K-mean clustering is used to detect disease on real-time leaf images. Once the detection has been made, the GLCM filter extracts its features. Features-based matching is implemented using an ANN approach for classification, and it attains 96% accuracy [14].
Jayamoorthy et al. (2017) compared K-means and Fuzzy C-Means with the suggested Spatial Fuzzy C Mean (SFCM) to identify plant diseases (FCM) and, likewise, kernel-based FCM (KFCM). The characteristics like the image of the sick leaf, color, and texture [25] were retrieved. The disorders were categorized using a neural network method. The recommended approach produced better results and suggested pesticides to combat the illnesses [16].
Kaur, N. (2021), emphasizing feature extraction, has considered GLCM, LBP, Gabor features, and SIFT. KNN, SVM, ANN, Logistic Regression, and Naive Bayes as a classifier to create an effective ensemble classifier. It has been found that the ensemble classification using the law's mask's hybrid features produced the best results in terms of recall, accuracy, and precision [15].
The approach for identifying and categorizing crop illness using machine learning integrated with digital image processing is presented in this work. With several image processing techniques for pre-processing, segmentation, and feature extraction, Ganatra et al. (2020) have suggested a prediction model for identifying and categorizing various plant leaf diseases. The disease is categorized into one classification using a variety of categorization techniques. On two independent datasets, they tested the suggested model [5].
S.No
|
Authors & year of publication
|
Plant
|
Disease Identified
|
Segmentation and Classification techniques used
|
Limitations
|
1. |
Youwen, T. et al. (2008)
|
Cucumber leaves
|
Powdery mildew and
Downy mildew
|
Segmented the leaf photos using statistical patterns and mathematical morphology. SVM classifier is used. SVM linear kernel is good.
|
Limited training sets of samples are appropriate for SVM. The linear kernel is reasonable if many features are obtained and the distribution is good.
|
2. |
Arivazhagan, S. et al. (2013)
|
Banana, tomato, Beans, Potato, Mango, Lemon, & Jackfruit
|
Early scorch, Brown spots,
Yellow spots,
A bacterial disease, Late scorch, and Fungal diseases
|
After color transformation, the infected leaf region is split into no equal patches, and the texture feature is extracted using the color co-occurrence approach. SVM classifier is used. The accuracy attained is 94.47%
|
The testing and training image size is only limited to 10 to 30. The computed value of texture properties is not specified.
|
3. |
H. Al-Hiary, S. et al. (2011)
|
Plant stem and leaves
|
Early Scorch, White Mold, Tiny Whiteness, Cottony Mold, Ashen Scorch, and Late Mold
|
Otsu threshold is selected in order to define the variable threshold value. K means clustering is used to segment the leaf image where 3 or 4 clusters produce the best result. The Color Co-Occurrence Method is used for feature extraction. A neural network classifier performs classification. The accuracy obtained is 94%
|
The Dataset descript-tion needs to be mentioned. Number of categories used for training and testing is not.
|
|
4. |
Sannakki, S. S. et al. (2013)
|
Grapes leaf image
|
Downy mildew and Powdery mildew.
|
Thresholding is applied for masking, and k means is used for segmenting images into 6 groups, and then extraction of feature is accomplished by GLCM. Feed forward back propagation neural network is used
|
As a whole, 33 images are used for training and testing.
|
5. |
Hossain, E. et al. (2019)
|
Plants
from Arkansas DB & redditplant leaf dataset
|
Alternaria Alternate, Leaf Spot, Anthracnose, Canker, and Bacterial Blight
|
Color segmentation is done using a k-nearest neighbor classifier with 3 neighbors.
KNN classifier is used.
|
It is highly inefficient to choose the "right" value K. here the authors used only 237 images, for huge data prediction will consume more time.
|
6. |
Samajpati, B. J. et al. (2016)
|
Apple fruit
|
Apple scab, apple blotch, and apple rot,
|
K means clustering is used for segmentation. LBP, CLBP, Local ternary pattern, and Gabor features approaches are used to extract a feature. A random forest classifier is used.
|
Training and testing data size is less. The Dataset description is missing. The local ternary pattern uses constant to threshold pixels, so the histogram will result in large range.
|
7. |
Pang, J. et al. (2011)
|
Maize leaf
|
Cercospora Leaf Spot, Brown Spots and Leaf Sheaths,
|
LTSRG segmentation algorithm.
|
Parameter specifications is not given
|
8. |
Ramesh, G. et al. (2020)
|
Four different categories from the plant image dataset
|
Alternaria,
Bacterial Blight disease, anthracnose,
and Cercospora Leaf Spot disease
|
K means clustering with value 5 is used. GLCM is for feature segmentation extraction. ANN is used for the classification
|
Parameters of the model must be used to control the underfit and overfit problems that frequently arise while training.
|
9. |
Jayamoorthy, S. et al. (2017)
|
Several plant species
|
Bacterial blight,
Footrot
|
They are comparing Spatial Fuzzy C Mean (SFCM) with other clustering methods to identify crop diseases Fuzzy C-Means (FCM), K-means, and Kernel-based FCM (KFCM), The Neural Network model is used.
|
In FCM, computation takes a long time. Due to its greater use of computational logic. Experimental results and accuracy need to be specified.
|
10. |
Kaur, N. et al. (2021)
|
Bell pepper, Potato, and Tomato
|
Early Bight,
Bacterial spot,
Curl Virus,
Target Spot,
Yellow leaf,
Mosaic Virus,
Septoria Leaf Spot
|
K means segmentation. GLCM, LBP, Law’s Texture mask, SIFT, and Gabor are used as feature extractors. Ensemble classifier with SVM, ANN, logistic, Naive Bayes, and KNN
|
The initial value of k needs to be specified. Parameters of the model must be used to control the underfit and overfit problems that frequently arise while training.
|
11. |
Ganatra, N. et al. (2020)
|
General Leaves
|
Early Bight,
Bacterial spot,
Curl Virus etc.,
|
Otsu's technique does the segments of the image. Support Vector Machine,
Artificial Neural Network,
Random Forest and K-Nearest Neighbor are used. The accuracy achieved is-73.38%
|
Neural network models have the propensity to overfit on smaller datasets. They must generalize successfully to new examples because they memorize the training data.
|
12. |
Zamani, A. S. et al. (2022).
|
Rice
|
Brown Spot, Leaf Smut, and Leaf Blight
|
Background noise is removed via the mean filter. The image quality is improved via histogram equalization. K-Means is used for segmentation and PCA for feature extraction. Then, images are categorized using methods like ID3, RBF-SVM, random forest, and SVM.
|
An RBF kernel SVM's complexity increases with the training set's size since the RBF kernel is not a parametric model. Here authors used a limited number of training and testing images.
|
13. |
Yang, X. et al. (2022).
|
15 spices
|
Leaf recognition
|
HSV color space segmentation method is used. SVM, BPNN, KNN, and BP-RBF are used for the classification
|
Backpropagation may be susceptible to erratic and inconsistent data. The training set of data significantly impacts how well backpropagation performs.
|
14. |
Badiger, M. et al. (2022).
|
Skin or leaf
|
Different types of diseases in skin and leaf
|
K-Means segmentation technique, SVM for classification of leaf and skin image.
|
If the target class in SVM overlaps, the algorithm could not perform as well.
|
15.
|
Ansari, A. S. et al. (2022).
|
Grape
|
Black rot,
Anthracnose, Leaf blight
|
Fuzzy c means the segmentation method used, HWT is used for feature extraction, PSO SVM, random forest, and BPNN algorithms are used for classification.
|
Limited number oftesting and training datasets. There is no certainty that an optimal solution will ever be found when using metaheuristics like PSO.
|
Through the literature survey, finally, the decision was to determine which basic segmentation would work best with which classifier on our CCDDB dataset. Basic segmentation like Global threshold, Adaptive Gaussian, Adaptive Mean, Otsu, Canny, Sobel, and K-Means is not applied to a single dataset. Basic Segmentation methods with classifiers have yet to be compared. The proposed paper used all of these segmentations and evaluated how well each classifier performed. In order to achieve better accuracy, this work determines which segmentation technique will work best with which classifier.