Agriculture has become a means to feed ever-growing populations. The plant has become a primary source of energy for humans as well as animals. It is also an important piece in the puzzle to solve the problem of global warming. Leaves of the various plants and herbs are helpful to mankind due to their medicinal attributes. In Asia and Africa where over 50% of population depends on agriculture production for employment, export earnings and food security (Anthonys and Wickramarachchi 2009; Singh and Kaur 2018)
It is assessed that 30% to 40% of crops are lost each year from production chain and in India, the annual estimated losses due to nematodes have been assessed to be about a 242.1 billion (Dhingra, Kumar and Joshi 2018)
which is having an impact on economical, social and ecological conditions. There are several diseases which affect plants These diseases can be broadly classified under parasitic or non parasitic causes as in Fig 1.
Fig 1: General structure of Types of plant diseases,(Dhingra et al. 2018).
Most diseases however generate some kind of manifestation in the visible spectrum. In most ofthe cases, the diagnosis about the disease is performed visually by humans. However this method is having disadvantages in accuracy, substantial inter and intra variability prone to various illusions (Anthonys and Wickramarachchi 2009)
. The foundation of quality assessment is basically dependent upon features of leaves such as its appearance, cracks, textures and surface where human alert can be easily fooled. In fact 60% to 70% of diseases appear on leaf only. So we have interest in the plant leaf rather than whole plant (Dhingra et al., 2018).
The review is restricted to the Rice plant and its diseases and the considerable research is also done on rice plant diseases. The commonly occurring rice diseases and their symptoms and causes are as in Table 1.
Table 1: Rice diseases, symptoms and cause.[Rice | Diseases and Pests, Description, Uses, Propagation (psu.edu)].
The important doctrine and standard Studies show that various image processing based machine learning and deep learning techniques are effective tools for identification and classification of plant diseases. The underlying principle and standard remain same, but computation and testing become accurate with the application of computer vision techniques.
The image analysis technique is used to distinguish the object from the background, there by isolating quantifiable information which is used in various decisions. Machine learning based, location and acknowledgment of plant infections give extensive information to recognize and treat the diseases. (Mallikarjuna, Guru and others 2011; Singh and Kaur 2018)
Many researchers have worked on the improvement strategies for the programmed recognition and classification of leaf infections, in light of high determination multi spectral, hyper spectral and stereo pictures. Disclosure of plant illnesses based on the symptoms of diseases showed on the plant leaves is the topic of present focus. The authors are searching for quick, affordable and precise technique to recognize plant diseases (Singh and Kaur 2018)
. The solution will help the agriculture domain through the automation of the plant disease detection and classification.
Classification is playing a key role in replacing the manual processes by the automatic process. The classification of the desired classes is automated using the various classification algorithms. The algorithm takes the physical features as input and gives the labeled object as output. Soil related features like, moisture, temperature, electrical conductivity pH, organic carbon, available nitrogen are considered as input data and soil type is considered as class label. Features are fed to the Linear Discriminate Analysis algorithm and are selected using the Boruta algorithm and its performance is found to be 96.3% indicated by scatter plot (Radhika and Madhavi Latha 2019)
Remote sensing is a new technology to sense things that are far away, from the ground, air and satellite. This Technique uses different imaging sensors and also a few non-imaging sensors to collect the data. Plant diseases reflect the various symptoms on various parts of the plants such as the leaf, stem, flower, etc
. These parameters can be sensed remotely to detect and monitor the health of the plant. Rice disease sheath blight was sensed remotely using Air Borne Data Registration in 2005. This has become the motivation for automating the identification and classification of plant diseases using technology like remote sensing, computer vision and machine learning. These techniques are helpful in large-scale farming to detect diseases on time. They also provide non-invasive, rapid, reliable, precise and accurate estimates of diseases which place a very crucial role in the monitoring of epidemics. (Gogoi et al. 2018)
Automation of agricultural processes is the recent advancements using computer vision and Neural Network based machine learning techniques. The dry red chilly images are acquired using image sensors and segmented using the Otsu thresholding algorithm to separate the chilly from its background. Morphological and color-based features are extracted. The dry chilly grading system is automated using self organizing map clustering algorithm (Shetty et al., 2020).
The affected part and the healthy part of the plant leaf is separated using thresholding segmentation techniques. This techniques will consider the task as two class problem and the threshold is calculated using local entropy, Otsu method (Anthonys and Wickramarachchi 2009
; Mallikarjuna et al., 2011; Phadikar and Goswami 2016)
, variable and global threshold values are calculated to segment an image into affected and non affected part (Guru, Mallikarjuna and Manjunath 2011
; Orillo et al., 2014).
For better segmentation results K-means clustering, region based segmentation are used in one or two stages (Arnal Barbedo 2013
; Guru et al., 2011; Hasan et al., 2019; Qiangqiang et al., 2015; Ramesh and Vydeki 2020; Singh and Kaur, 2018)
The performance of the segmentation technique is evaluated by few researchers using error rate, over lapping, under and over segmentation, Dice similarity, Precision and Recall (Anthonys and Wickramarachchi 2009
; Qiangqiang et al., 2015).
The classification methods used by many researchers takes the features of the image and label of the classes as input and get training further for testing phase. The following are the features extracted in many researches.
Shape features are extracted from the segmented image using neighborhood technique and Blob analysis (Guru et al., 2011; Mallikarjuna et al., 2011; Orillo et al., 2014; Pinki, Khatun and Islam 2017)
. Texture features of first and second order are extracted by contrast stretching transformation (Anthonys and Wickramarachchi 2009; Mallikarjuna et al., 2011; Pinki et al., 2017)
and gray level co-occurrence method (Arnal Barbedo 2013
; Hasan et al., 2019).
Fractal descriptors are derived from fractal dimension. (Asfarian et al., 2013; Devi and Muthukannan 2014)
Color features are extracted using color components and by their histograms (Arnal Barbedo 2013; Khirade and Patil 2015; Ramesh and Vydeki 2020)
. The contrast energy, energy relation and homogeneity are also used for classification (Phadikar and Goswami 2016)
Automation is achieved by many researchers using the machine learning for classification. Considerable accuracy is achieved using probabilistic neural network, fuzzy classification model, back propagation neural network, random forest algorithm, support vector machine, decision tree and convolution neural network. Performance of the classification models are evaluated using confusion matrix, performance plot, training state plot, accuracy, cross validation are addressed by most of the researchers (Devi and Muthukannan 2014
; Dhingra et al., 2018; Guru et al., 2011; Hasan et al., 2019; Khirade and Patil 2015; Singh and Kaur, 2018)
. The correlation between the performance of classification method and precision of the segmentation technique is mostly unaddressed by the researchers.
The proposed work is carried out in the Department of Electronics and Communication research center at Ballari Institute of Technology and Management, Ballari, Karnataka during the period from June 2021 to August 2022 in collaboration with the expert from the department of plant pathology, university of Agricultural Science, Dharwad, Karnataka. The samples required for the work are taken from machine learning repository [https://www.kaggle.com/datasets/vbookshelf/rice-leaf-diseases].
The methodology used by most of the researchers includes the following steps in identifying and classifying the plant diseases. Image acquisition Image pre-processing, Segmentation, Feature extraction, classification and identification and Performance evaluation are as per the sequence given in below Fig 2.
Fig 2: Methodology for plant disease detection,(Khirade and Patil 2015).
The input data required are images of different diseases, collected from the paddy fields using the digital camera and digitized into [M] rows and [N] columns (Singh and Singh 2015)
directly in the RGB color model using digital cameras and stored in the computer memory for further processing (Mallikarjuna et al., 2011; Orillo et al., 2014).
These images are transformed into smaller dimensions (Arnal Barbedo 2013; Ramesh and Vydeki 2020)
. Images are collected in controlled environmental factors by using a controlled light module box (Khirade and Patil 2015)
Image Pre-processing and image segmentation
Making an image ready for segmentation by applying image enhancement, blob analysis for noise reduction, contrast adjustment and image intensity adjustment (Khirade and Patil 2015)
Converting images from RGB color models into HSV, Gray images (Mallikarjuna et al., 2011; Singh and Singh 2015)
. Image smoothing and linear filtering (Dhingra et al., 2018)
are performed as a pre-processing steps.
The next step is to segment an image into a region of interest and background or diseased part and healthy part of the leaf. It is found in literature that, segmentation is performed by Local entropy, Otsu thresholding methods (Mallikarjuna et al., 2011),
Region growing segmentation and Mean shift segmentation (Pinki et al., 2017).
Histogram equalization using two bins and masking with the original image and segmented image. Segmentation using color model boundary and edge detection using the Sobel operator. (Arnal Barbedo 2013)
K-means clustering (Phadikar and Goswami 2016; Ramesh and Vydeki 2020; Singh and Kaur 2018)
. Active contour models and super pixels segmentation (Qiangqiang et al., 2015)
are also used for segmentation.
Features extraction in image
The next step is to extract the characteristics from the segmented image. Features extracted from acquired images are, statistical features using, color co-occurrence matrix or gray image, shape features using blob analysis (Arnal Barbedo 2013
; Hasan et al., 2019; Khirade and Patil 2015; Phadikar and Goswami 2016)
and Fourier dimension and descriptors (Anthonys and Wickramarachchi 2009; Narmadha and Arulvadivu 2017)
Detection and classification of plant disease
These features are fed to the classification model and the methods used in classification are, the production rule based method through serial interviews (Mallikarjuna et al., 2011; Orillo et al., 2014).
Features mapped to the disease will decide the membership function (Singh and Singh 2015)
and the machine learning-based supervised and unsupervised algorithms are used to identify the disease type. The model preparation is done in two phases. one is training phase and the other is testing phase. In the training phase, the extracted features are prepared with the name of the disease as the label. These are given to the models like Support vector machine, Random forest classifier (Hasan et al., 2019),
probabilistic neural network (Devi and Muthukannan 2014; Khirade and Patil 2015; Narmadha and Arulvadivu 2017)
, linear discriminate analysis and Back propagation algorithm (Arnal Barbedo 2013; Ramesh and Vydeki 2020)
and neural network models. The crucial steps in the above methodology are segmentation and classification. Accuracy of segmentation is measured using Jaccard similarity. Over and under segmentation is measured through (Mallikarjuna et al., 2011; Pinki et al., 2017; Singh and Singh 2015)
structural components, normalized cross-correlation, peak signal to noise ratio based on difference in the image pixel value (Guru et al., 2011).
Confusion matrix, accuracy, sensitivity, specificity, Dice’s coefficient, precision using true/false positive and negative rates are used as performance evaluating metrics for classification (Dhingra et al., 2018; Kaur and Oberoi 2020; Khirade and Patil 2015; Zhang, Yan and Hou 2018)
The literature review has revealed that the automation of Rice plant health monitoring and classification of diseases based on the features of the affected part is an open-ended problem. The techniques supporting the development of solutions are from image processing, pattern recognition and machine learning domains.
Various segmentation techniques of image processing are used to separate the healthy part and affected parts. Based on the precision of segmentation; the techniques are classified as primary, secondary and higher segmentation for this work. Thresholding, local entropy and Otsu segmentations are treated as primary techniques. Region growing clustering, mean shifting and super pixel segmentation as secondary techniques. K means, Active contour are considered as higher level techniques.
The segmentation output of disease affected paddy leaves are obtained using K-means segmentation and Otsu thresholding method then the disease part and the normal parts of the leaves are visually inspected and termed as properly and poorly segmented outputs as shown in Fig 3(a) and 3(b).
Fig 3(a): Properly segmented image.
Fig 3(b): poorly segmented image.
The statistical features say contrast, energy, homogeneity, mean, standard deviation, entropy, root mean square(RMS), variance, smoothness and inverse difference moment (IDM) are extracted from segmented images and used as input for multi class support vector machine classification algorithm.
The classification algorithm support vector machine(SVM) results were analyzed and the classification accuracy was found to be 76.60%, for properly segmented images against the 58.30% accuracy for the poorly segmented images (Fig 4).
Fig 4: Classification accuracy.