Machine Learning and Image Processing Techniques for Rice Disease Detection: A Critical Analysis

Abstract

Early rice disease detection is vital in preventing damage to agricultural product output and quantity in the agricultural field. Manual observations of rice diseases are tedious, costly, and time-consuming, especially when classifying disease patterns and color while dealing with non-native diseases. Hence, image processing and Machine Learning (ML) techniques are used to detect rice disease early and within a relatively brief time period. This article aims to demonstrate the performance of different ML algorithms in rice disease detection through image processing. We critically examined different techniques, and the outcomes of previous research have been reviewed to compare the performance of rice disease classifications. Performance has been evaluated based on the criteria of feature extraction, clustering, segmentation, noise reduction, and level of accuracy of disease detection techniques. This paper also showcases various algorithms for different datasets in terms of training methods, image preprocessing with clustering and filtering criteria, and testing with reliable outcomes. Through this review, we provide valuable insights into the current state of ML-based approaches for the early detection of rice diseases, and assist future research and improvement. In addition, we discuss several challenges that must be overcome in order to achieve effective identification of rice diseases.

1. Introduction

The need for agricultural development is expanding steadily with each passing day due to the growing global population. Meanwhile, 90% of people across the globe are dependent on agriculture. Farmers’ hard work and commitment are recognized as the driving force behind the production of food that feeds 80% of the world’s population [1]. Insect damage to agricultural plants causes a loss in productivity, which negatively influences our economy. The development and growth of edible plants are significantly affected by plant diseases [2]. The agriculture industry is currently dealing with serious issues of plant diseases that diminish harvest quality and output. The agricultural industry is experiencing disruptions due to a lack of trained personnel in the farming sector, poor information on fertilizer application, and ignorance regarding diseases and insects [3]. Considering agricultural impacts on global nutrition, rice holds a position of utmost importance as a crop that is extensively grown and harvested worldwide. Leaves are the most sensitive parts of the plant where disease symptoms first appear [4]. Early disease detection and classification are very helpful for agricultural disease management. In traditional disease observation, specialists have to physically check the crop areas to watch for viruses, which is tedious and time-consuming [5]. Therefore, a variety of approaches have been used as semi-automatic and autonomous disease detection systems for plants. Thus far, automated disease detection systems have been shown to be faster, more cost-effective, and more accurate than the farmers’ human observation method [6]. In many countries, agriculture has become the most important industry, contributing to the economic development of the country. Considering the significance of farming, farmers make conscious decisions regarding the selection of their crops, plows, and associated chemicals to encourage plant growth in a short amount of time [7]. Crop diseases create some environmental disruption directly and indirectly. These diseases, which are spreading all over the world, are wreaking havoc on plant health and the economy by drastically decreasing the amount of food produced [8,9]. Rice diseases of various kinds, such as leaf smut, rice bacterial blight, rice blast, and rice brown spot, are seen in paddy fields and are caused by bacterial infection, viral infection, and fungal effects [10]. Each disease exhibits distinct or varying characteristics, such as the colors, shapes, and patterns of the affected portions of the leaves. Four common diseases that affect rice plants are Brown spot, Leaf Smut, Bacterial Blight, and Rice Blast [10].

Table 1. Some rice diseases in agricultural fields and their causes and symptoms.

S.N.	Name	Image	Causes	Symptoms
1.	Bacterial Blight		Bacteria are the cause of paddy bacterial blight, a disease. The scientific name of this bacterial disease is Xanthomonas oryzae.	The symptoms of this disease manifest in various ways, including the appearance of water-soaked stripes on leaf edges, yellow or white stripes on leaf edges, grayish-colored lesions, shriveling and rolling up of plants, yellowing of leaves, stunted growth, plant death, and yellowing of the youngest leaf on the plant.
2.	Brown Spot		Brown spot is another fungal disease having the scientific name Cochliobolus miyabeanus.	The appearance of brown-spot-affected rice leaves is characterized by round dark brown circles. The circles have a diameter of 0.5 or smaller. They can be easily identified from a distance by their brownish scorched appearance.
3.	Rice Blast		Blast, a fungal disease that affects rice plants, is caused by a fungus known as Magnaportheoryzae.	All the parts above the ground can be affected by rice blasts. The primary symptom is elliptical, diamond-shaped, or spindle-shaped spots with grey or white centers and brown margins.
4.	Leaf Smut		The fungus Ustilaginoidea virens causes leaf smut.	It causes small black spots to appear on the leaves. A group of pathogens infect the rice plant during the flowering stage and cause grain chalkiness. The spots are slightly raised on the leaf, and the leaf looks as if it were sprinkled with black pepper. Tips of the leaf may often die and become black.

shows different kinds of paddy diseases, including the colors and symptoms of leaves.

If rice is cultivated on soils lacking silica, brown patches typically become worse [11,12]. However, visual examination of the leaf color combinations and crown architecture is still the primary method used in conventional techniques for crop disease investigations. Different image-processing strategies have been utilized for disease management in the agricultural field [13]. It is quite challenging to detect rice disease at an early stage based on colors, spots, or streaks on leaves or stems in image processing techniques [14]. Due to the diversity of plants, varieties of plants also exhibit diverse disease features, which add many complications to the classifications of plant disease. In the meantime, numerous studies have concentrated on the machine-learning-based diagnosis of plant viruses [15,16,17].

1.1. General Structure

The basic steps of the rice plant disease detection system are:

○

Image acquisition;

○

Image preprocessing;

○

Image segmentation;

○

Feature extraction;

○

Classification.

1.1.1. Image Acquisition

Figure 1 illustrates the fundamental framework for categorizing diseases affecting plant leaves. In image processing, the process of obtaining a photo from a given source, which may involve manual image capture or accessing a dataset, is commonly referred to as image retrieval. Researchers captured images using high-resolution cameras and afterwards resized them to a specific number of pixels for their studies [18]. In addition to other methods, the UC Irvine Machine Learning Repository [19] and ImageNet [20] datasets are commonly used for image acquisition.

1.1.2. Image Preprocessing

The process typically involves resizing, cropping, and eliminating noise from the original image to achieve the desired output. The main objective of the preprocessing phase is to enhance specific image characteristics essential for subsequent processing. This phase entails removing some undesirable features from the provided image. The image is divided into clusters in the second phase of preprocessing [21]. In order to reduce the image processing time, certain nonessential portions of the image, such as the background and insignificant segments, are removed [22].

Image enhancement plays a crucial role in disease detection. It refers to the process of improving the quality, clarity, and interpretability of an image, making it easier for professionals to detect and diagnose diseases. In summary, image enhancement is crucial for rice leaf disease detection as it enhances the visibility of disease symptoms, aids in differentiation from healthy leaves, reduces noise, standardizes images, facilitates quantitative analysis, and supports automation. By improving the quality and interpretability of rice leaf images, image enhancement contributes to effective disease management, higher crop yields, and sustainable agriculture.

1.1.3. Segmentation

Image segmentation is employed to partition images into distinct segments for analysis. This process involves transforming an image into a more comprehensible format, thereby enabling easier analysis. In this instance, an image can be segmented to identify different regions based on the specific feature of interest [23,24]. The K-Means Clustering [8], contours [9], Otsu segmentation method [19], and region segmentation [25] are examples of segmentation techniques.

In addition, thanks to recent breakthroughs in deep learning, Convolutional Neural Networks (CNNs) have become more popular for object identification applications. There is no longer any need for explicit segmentation because of CNNs ability to immediately learn and extract characteristics from pictures [10]. Several implementations of this strategy have shown considerable promise.

1.1.4. Feature Extraction

The method for feature extraction extracts attributes from segmented data based on colors, textures, and shapes [10]. Area, angle, and axis are shape-based properties [26]. Figure 2 shows a CNN feature extraction framework for HSI data, which includes convolution layers, pooling layers, and a fully dense layer.

1.1.5. Classification

Identification of plant disease systems relies heavily on classification. It is referred to as the procedure of categorizing rice leaf images based on diseases that have been identified. The categorization techniques can be broadly classified into two types, namely, supervised and unsupervised methods. Classifying images involves utilizing a highly skilled classifier that has been thoroughly trained. The unsupervised order uses the characteristics of pixels to group them, which is known as clustering [27]. To carry out the classification process, supervised techniques like K-Nearest Neighbor, Logistic Regression, Naive Bayes, and Decision Tree were utilized [28,29]. In addition, the artificial neural network is an emerging classification method.

Numerous machine learning (ML) and image processing algorithms have been developed for disease identification in rice plants. Accuracy in diagnosing plant diseases using ML algorithms depends on three procedures: feature segment, extraction of features, and verification method. Deep learning techniques have produced favorable results for picture classification. The convolution neural leaf disease diagnosis model can differentiate disease-affected leaves with symptoms of major diseases. Since CNN does not require the time-consuming preprocessing of input images or the creation of customization options, CNN is chosen over traditional methods in several applications. This method has a faster number of iterations and provides better performance [30]. The infected parts of the images are analyzed as segments of leaf images in these techniques. The proper feature extractions and classification of diseases are applied to recognize the infected leaf. The system will be accurate if it detects the infected leaf through image processing and classification [31].

The ML method is incorporated to identify the rice leaf disease through classifying sets of pixels or vectors. “Supervised” and “unsupervised” identification techniques are used as common types of image classification in disease identifications [32]. Sometimes, spectrum analysis and textural properties are used to obtain more reliable outcomes in agricultural disease detections through image processing [33].

2. Review Planning

This systematic examination aims to give an overview of different machine learning and deep neural network techniques utilized in detecting paddy plant diseases. Figure 3 illustrates the comprehensive flow of our research methodology. The review methodology involves the following steps:

• Data Gathering
• Search Approach

This literature review focuses on examining digital scientific databases, and therefore, no forms of gray paper were included. The study is founded on the premise that the majority of research containing relevant results, often found within the gray literature, is typically acknowledged or referenced in previously published scientific papers.

In this literature analysis, databases were employed, namely Science Direct, ACM Library (Association for Computing Machinery), Scopus, Springer, and IEEE Xplore. The reason for selecting these databases is that they offer a vast quantity of high-impact scientific journal articles and conference proceedings, which provide extensive coverage of the research field regarding plant diseases and machine learning via CNNs.

Inclusion procedures: To identify papers that meet the desired criteria, the inclusion criteria involved the initial selection of titles and abstracts, followed by removing duplicate papers.

Exclusion procedures: The study excluded papers that did not focus on detecting and grouping rice or other plant leaf diseases using deep learning/CNN. However, we kept some papers related to image enhancement in case they could improve the detection or classification performance.

Data Analysis: To conduct data analysis for the review, approximately 100 papers were selected based on their relevance, and the following points were considered.

Publication Year: A crucial aspect of analyzing the increasing interest in recognizing paddy diseases utilizing CNN/deep learning is to determine the year of publication, as this topic has garnered significant attention from researchers in recent decades.

Objective: The study comprised a variety of tasks, such as lesion and discoloration detection, segmentation, and classification, targeted at identifying and diagnosing various rice diseases.

Deep Learning (DL) Structure: Various plant disease identification and pest tracking techniques have utilized deep learning architectures, such as Deep Neural Networks, Convolution Neural Networks (CNN), and RNN.

3. Literature Review

The ML method has been applied for early disease detection through image processing. Gayathri et al. proposed a structure that automatically detects rice leaf diseases using image processing techniques. Image preprocessing has been included in the methods for leaf disease detection. The features are extracted using the Gray Level Co-occurrence Matrix, Discrete Wavelet Transform, and scale-invariant feature transformation techniques. The extracted features were processed through different classifiers, such as K-Nearest Neighbors and neural networks (NN) (KNN), classification SVM, and Naive Bayes, to differentiate healthy and unhealthy plants. The results showed that the multi-class SVM performed better than other techniques, with a precision of 98.63% [29].

Mittal et al. proposed an algorithm with icon-centric information-gathering architecture. This model is more accessible to farmers, allowing them to diagnose crop diseases immediately using pattern recognition and digital image processing rather than having to wait for specialists to come to the fields and diagnose the disorders. Results from the experiments showed that the methodology is more straightforward and more cost-effective than other methods of digital image processing. The image processing part presented the results based on training 25 samples for each type of disease. Better performance would be obtained by increasing the number of samples of affected leaves [31].

In order to minimize the noise effects for disease detection, researchers used a histogram as an established framework for analyzing the images. This system allows for the earlier-than-usual identification of diseases and the implementation of preventative measures to lessen productivity loss. The plant’s image was captured first, and then it was filtered for disease identification. The leaf image was modified from RGB to grayscale, and then researchers used a histogram to extract the features using MATLAB methods [30]. The images obtained were used to classify and evaluate the diseases. A collaborative treatment was developed to identify the various kinds of paddy diseases, including bacterial leaf blight (BLB), leaf streak, sheath blight, brown patches on leaves, false smut, and RLB diseases caused by both fungi and bacteria. The author of this paper suggested an automated system. This system included image acquisition, feature extraction, leaf verification, and preprocessing in image processing techniques. Features of standard deviation, mean, difference, and relationship of the color, statistical, and textural aspects were extracted using Principal Component Analysis (PCA) and Grid-based Color Moment (GBCM) [34].

SVM algorithms were used to classify the diseases based on leaf disorders. War et al. proposed creating an image database using familiar farm photos that were immediately captured in discrete numerical values. For the original grayscale conversion and the updated grayscale conversions, this technique achieved efficiency with 72.730% and 90.2%, respectively. Image processing, as well as machine learning (ML) techniques, have been applied to identify viruses in rice plants [13].

Liang et al. proposed a CNN-based technique to deal with 2906 progressive samples from public datasets for evaluating and training the system. Different rice diseases are detected more accurately using CNN techniques for feature extractions. Identity models were fused with domain knowledge related to rice diseases to achieve computerized rice disease diagnosis. In addition, qualitative and quantitative analyses were conducted in their research to evaluate the performance of the CNN-based technique. These results showed that the deep-level features extracted by convolutional neural network (CNN) techniques are more effective and discrete than the conventional handmade features like Haar Wavelet Transform (WT) and local-binary pattern (LBP) histogram. In addition, quantitative evaluation demonstrates that CNN with SVM and CNN with Softmax increase the reliability in agricultural disease identification [35].

Lu et al. used 500 images of leaves as a dataset. This dataset is captured from the rice plants. CNN algorithms classify the datasets on 10 disease leaves. According to the results, the CNNs-based model proposed in the study achieved an accuracy of approximately 95.48% when the 10-fold cross-validation strategy was employed. Compared to traditional machine learning models, this accuracy is far higher. The model is validated for the diagnosis of rice diseases, demonstrating the viability and efficiency of the suggested approach [36].

Mique and Palaoag proposed a model to identify rice diseases and insects through image processing and CNN techniques. This model included insect identifications, farmers’ awareness about pests and illnesses, and agricultural field management criteria. An algorithm was proposed for identifying rice insects and illnesses using image recognition and CNN in the agricultural field. This model primarily achieved the goal with a recognition accuracy of 90.90%. A lower cross-entropy value suggests that the trained scheme makes predictions earlier or classifies images more accurately. Farmers can control and manage rice insects through this cutting-edge application [37].

Pothen and Pai suggested a method for detecting Leaf Spot, Bacterial Blight, and Brown Spots. HOG is used to classify features derived from the Support Vector Machine and LBP. Otsu’s segmentation method was used to differentiate the infected and healthy sections. Otsu’s method of segmentation determines the most suitable value for the threshold. To segment an image into sections, the Otsu approach groups pixels representing the same items in an image [19].

Nidhis et al. suggested an image-processing-based system to recognize Leaf Blast, Bacterial Blight, and Brown Spot. This article used k-means clustering to distinguish the damaged and undamaged parts of the leaf. The point feature matching strategy was used to detect disease. Size, color, proximity, and centroids categorize different diseases [38].

Zhou et al. offered an approach to identify and classify diseases, for example, bacterial blight, sheath blight, and rice blasts. The threshold approach of Otsu was used to reduce background interference by detecting the necessary region. The FCM km technique was used to determine the best k values. FCMKM and Faster R-CNN were combined to recognize rice diseases. Leaf blight detection, bacterial blight, and rice blast accuracy levels were 97.53%, 98.26%, and 96.71%, respectively [39].

Shreekanth et al. offered an approach for identifying Leaf Blight, Brown Spot, and Leaf Blast in rice leaves. Otsu’s approach was used to partition the image. For classification, texture features and wavelets are used. A feed-forward neural network (FFNN) is used for classification. Classification accuracy was 83.3% for three kinds of diseases and 100% for two kinds of diseases [40].

The goal of the work of Kawcher et al. is to identify diseases like rice Brown Spot, Bacterial Blight, and Leaf Smut using four different machine learning (ML) algorithms like Naive Bayes (supervised ML algorithm), Logistic Regression, K-Nearest Neighbor (k-NN), and Decision Tree. For 10-fold cross-validation, the model’s disease detection accuracy was 97%. A correlation-based approach to feature selection was used in this paper to extract attributes (features) [41].

Bakar et al. presented a method for identifying paddy Leaf Blast disease, which also classifies disease severity into three phases: spreading, worst stage, and infection. The HSV color space is used in this piece. To locate the region of interest, a multi-level thresholding technique was used. Color and shape characteristics were extracted. This approach is not suitable for identifying other diseases with similar symptoms [25].

Ramesh and Vydeki developed a method that employs a machine learning (ML) technique to detect paddy Leaf Blast disease early. The K-means clustering technique was used to divide the damaged part of the image from the healthy portion. To differentiate between diseased and healthy leaves, texture features and statistics are used. The ANN algorithm was used to classify pictures of good and diseased leaves [42].

Zhang et al. presented a technique to identify three different variations of rice blast damage, namely, acute type, white type, and chronic type. For segmentation, Otsu’s method was used. Morphology and color characteristics are used in classification. To classify the three different forms of paddy blast lesions, a classifier with a Support Vector Machine (SVM) and a Radial Basis Function (RBF) is used. The overall classification performance obtained by five-fold cross-validation is 95.6%. The segmentation process and feature methods used in this study could be applied to other crops [43].

Islam et al. presented an approach for detecting and classifying diseases such as Rice Leaf Blast, Brown Spot, and Bacterial Blight. It employs a percent of the RGB as a parameter to classify diseases using the Naive Bayes. It makes no mention of segmentation techniques. The validity of this approach is greater than 89 percent for Rice Blast disease and greater than 90 percent for Bacterial Blight and Rice Brown Spot disease [22].

Ghyar and Birajdar introduce a system for detecting Brown Spot Rice Leaf Blast disease automatically using color and texture descriptors. Three features are used for classification: texture descriptors, area, and color moments. The feature selection process employs a genetic algorithm. This system employs ANN classifiers and SVM. SVM has an accuracy of 92.5%, while ANN has an accuracy of 87.5% [44].

Hua Yang and others used improved YOLO v3 algorithm for detection of rice pest and diseases. The improvement in accuracy was 2.7%. The speed also improved to a notable point [45]. YOLO v3 has vast use in different dimensions, like in the detection of construction cracks [46].

Joshi et al. presented a technique for Brown Spot, Rice Leaf Blast, Sheath Rot, and Bacterial Blight. Objects were classified based on their shape and color characteristics. This system employs MDC (Minimum Distance Classifier) and K-NN. K-NN obtained 87.02% classification accuracy, while MDC obtained 89.23% accuracy [47].

4. Comparative Analysis

4.1. Image Processing Methods Employed to Identify Rice Disease

This section describes various image processing methods that researchers have employed to identify rice diseases in their studies, considering factors such as segmentation methodologies, segmentation type, dataset size, features retrieved, and picture background. We reviewed 19 studies on rice leaf and seedling disease conducted throughout the past eight years; the summary is presented in

Table 2. A summary of various methods for image processing used in rice disease detection.

Ref.	Techniques Employed	Extracted Features	Segmentation Type	Size of Dataset
[42]	Distance transformation	Shape and Texture	Watershed	2000 microscopic images
[47]	YCbCr color space	Color and Shape	Clustering	115 images
[46]	K-means	Area, GLCM-based texture descriptors, and color phases	Clustering	Not defined
[11]	K-means	Shape and color	Clustering	Not defined
[48]	K-means	Color and Shape	Clustering	Not defined
[44]	K-means	Color, shape, and texture	Clustering	Not defined.
[22]	Not Specified	RGB values	Not Specified	60 images
[19]	Otsu’s method	Radial hue distribution	Thresholding	1000 images
[49]	Otsu’s Method	The statistical properties and fragments, specifically the area, were analyzed in the study.	Thresholding	134 images
[45]	Otsu’s Method	Morphology and color	Thresholding	90 images
[36]	Pixel-based	Color and Structure	Multi-Level Thresholding	Not defined
[43]	Not Specified	Correlation Based Feature	Not Specified	480 images
[41]	K-means and Otsu method	Color and Texture	Clustering and Thresholding	3010 images
[50]	K-means	Color, size, proximity, and centroids	Clustering	Not defined
[51]	K-means and Otsu method	Color and Shape	Clustering and Thresholding	500 sample images
[52]	Not Specified	Features of morphology and color	Thresholding	700 images
[53]	Multiscale transform	Color	Fractal Descriptors	40 images
[54]	K-means, Otsu technique, and Fermi energy based	Shape, color, and position	Thresholding and Clustering	500 images
[55]	K-means	Texture and Statistical	Clustering	300 images
[56]	K-means, Jaya algorithm	Color and Texture	clustering	650 images

.

4.2. Methods of Segmentation for Detecting Rice Plant Diseases

There are several methods of segmentation that can be used for detecting rice plant diseases. Segmentation refers to the task of partitioning an image into meaningful and relevant regions or parts. The primary objective is to identify and isolate the specific parts of the image that are pertinent to the identification of diseases in paddy plants.

Table 3. An assessment of different segmentation methods for detecting rice diseases.

Ref.	Algorithm Used	Merits	Demerits
[57]	OTSU	The efficiency and accuracy of segmentation were excellent.	Isolated noises and holes remain largely in the picture after segmentation.
[58]	Panicle SEG	Enhanced segmentation accuracy. The execution speed has been increased.	Could not be used in a variety of field environments.
[59]	K-means clustering (KMC) algorithm	The disease was automatically identified. The sign color, illumination, and leaf color were separated using the diverse color channel.	The plant’s pigment was not extracted properly, and disease types were not distinguished.

provides a list of the segmentation techniques used in the diagnosis of rice diseases.

4.3. Feature Extraction for Rice Leaf Diseases Detection

The process of disease identification in rice involved the extraction of key features such as color, shape, and texture from leaf images, as presented in

Table 4. Comparison of various approaches of recognition and feature extraction used in leaf diseases diagnosis.

Ref.	Feature Extraction	Recognition	Advantages	Disadvantages
[38]	Edges, lines, and corners.	CNN	The stochastic pooling layer improves the development of the CNN.	The drawback is that not only is judgment easily inaccurate, but efficiency is also lacking.
[39]	Edges, lines, and corners.	CNN	Trying to improve the accuracy	Used limited dataset
[50]	color, texture, and size	SVM	More accuracy in disease identification	Need more features based on Leaf smut
[62]	Shape, color features	SVM	Successfully categorized four different types of rice diseases	The lowest level of accuracy observed in comparison to the others.

. Red–Green–Blue (RGB) plays an important role when differentiating one image from another in image processing [32]. Principal Component Analysis (PCA) is utilized for feature extraction on an unsupervised dataset of disease-affected leaves [60]. It is a technique for determining distinctive elements that describe an image using three features, including the picture’s color, texture, and size [61].

4.4. Comparative Analysis of ML Method in Disease Detections

This section presents a comprehensive survey of diverse machine learning algorithms that have been employed for the purpose of detecting and classifying diseases in rice. Our survey considers a range of parameters, including the author’s name, the techniques used, identified plant disease, dataset, dimensions, and accuracy. This survey aims to provide readers with a comprehensive review of the present state of study in this area.

Several Machine Learning and image processing approaches used in the identification and classification of leaf diseases are compared and delineated in lower

Table 5. A comparison of the different initiatives used in the diagnosis of leaf diseases.

Ref.	Technique Used	Disease Identified Plant	Dataset	Dimensions	Accuracy
[38]	Deep convolutional neural networks	Rice Leaf	500 images	5760 × 3840	95.48%
[39]	CNN	Rice Leaf	Gather data on the different rice pests from the Department of Agriculture	420 × 450	90.9%
[50]	SVM KNN	Rice Leaf	NIKON D90 Dataset	2848 × 4288	93.33%
[63]	Deep CNN-centered classification	It was detected that rice plants have pests and diseases	In real life, 1426 images were collected.	222 × 222	95%
[10]	Unified Modelling Language, Waterfall Paradigm and ES for RPD2 application	48 signs of the rice plants and 8 different disorders were identified	Rice images	1024 × 1024	87.5%
[64]	INC-VGGN	500 rice images and 466 images of maize	PlantVillage	450 × 470	91.83
[65]	PCA and NN	Rice Blast	The Zen Z3 camera was utilized to capture the image.	4000 × 3000	95.83%
[66]	PSO	Rice blast, SheathRot, Leafbrown spot, BB	Benchmark datasets	750 × 850	84.02%
[47]	KNN, J48 Naïve Bayes	Rice Leaf	40 images of rice leaf	Variable size	97%
[67]	CNN	Rice Leaf	Agriculture field	256 × 256	96.3%
[68]	CNN	Rice Leaf	3000 Rice images	Variable size	90%
[15]	Fuzzy neural network	Rice Leaf	45 images	450 × 450	93.33%
[69]	YOLOv3 tiny and YOLOv4 tiny	Rice Leaf	762 images	550 × 550	97.36%
[70]	CNNIR-OWELM	Rice Leaf	115 images	650 × 650	94.2%
[71]	DL model, SVM	Rice Leaf	5932	-	98.38%

.

According to Table 5, Prajapati and Shah et al. used SVM and KNN with a dataset of 2848 × 4288, where 93.33% accuracy was achieved in disease detections. Bakar et al. offered a model for image processing and recognition through thresholding. For image verification, some important steps are involved, such as preprocessing, image segmentation, and analysis of the leaf image using some algorithms. In this study, the researcher used a Hue Saturation Value (HSV). The image recognition was then applied to a leaf dataset on the basis of the spread stage, damaged areas, etc. [25]. AlexNet and GoogleNet, used for image processing, achieved 97.28% accuracy in agricultural disease detections and classifications [72]. Ahmed and Shahidi et al. (2019) attained 97% accuracy when using the KNN, J48 Naïve Bayes method with 40 images of rice leaf in image processing. Jain et al. (2022) used the YOLOv3 tiny and the YOLOv4 tiny method with 762 images, which achieved 97.36% validation accuracy in disease detections.

An SVM classifier uses genetic algorithm (GA) to select the desired qualities and limitations. The study’s findings revealed that the proposed approach achieved a validation accuracy of 87.9% and a positive predictive value of 91.8% in effectively differentiating between healthy and infected seedlings [57]. CNN and deep learning algorithms were used for leaf disease detection. For testing or verification, researchers used 1426 leaf images as a dataset. The dimension of the leaf’s dataset was 222 × 222 in image processing. This dataset consists of Bacteria leaf blight, Brown Spot, Brown Plant hopper, False Smut, and Stemberg images. CNN recognized the accuracy as 95% in disease detections and classifications [68].

The graph below Figure 4 summarizes the various Deep Learning models used to detect various types of rice plant diseases.

5. Discussion

Numerous methods are available to detect rice diseases, including edge detection, water separation, clustering, saliency, active contour, and thresholding. Regardless of the approach used, the process of detection in all methods involves five main steps, which include image acquisition, image preprocessing, segmentation, feature extraction, and classification. In agricultural sciences, the pre-treatment level may be utilized to (1) identify diseases of rice plant leaves; (2) identify sick parts of rice plants; (3) identify the shape and hue of rice crops and diseased areas; and (4) discover a remedy for rice diseases. Segmentation, which involves dividing the image into regions of interest, is the crucial technique for extracting image features. Texture-, image shape-, and color-related properties are among the essential image features that characterize the image. Several research articles have proposed automatic detection methods for leaf disease, and have discussed various segmentation and feature extraction techniques, including thresholding, edge-based, region-based, wavelet transformation, and Gabor filtering. In order to remove the background, suggestions are made for segmentation and background removal techniques. Color segmentation is performed using green pixel masks, followed by the application of Otsu thresholding methods to diseased images [73]. CNN utilizes characteristics, such as similarities, contradictions, energy, and clusters, to define shadows and their branches, enabling the determination of the forms of disease.

Machine learning (ML) and deep learning techniques have been incorporated with image processing for rice disease detection at early stages of diseases. Researchers have taken into account several factors to suggest the best techniques, including simplicity and reliability of the system, identification speed and accuracy, and appropriate categorization. The classification algorithms include the steps of image acquisition, image recognition, and feature extraction for disease diagnosis employing image processing techniques. Several researchers have changed images from the RGB values to the grayscale image to process images more efficiently.

Through image processing, segmentation of the necessary part is achieved and robustness of data during training is also increased. Filtering and ML process play important roles in these cases as given in Table 2.

Different algorithms are used in segmentation for detection of diseases; however, this also has some drawbacks. The contrast of merits and demerits are detailed in Table 3.

Feature extraction plays a crucial role in improving the performance of machine learning (ML) models by transforming raw input data into a more suitable representation. As stated in Table 4, various types of feature extraction implies various goals of detection. Depending on the regions of interest, features like shapes, edges, lines, colors, textures are extracted.

The reliability of the various techniques has been displayed using a graph in Figure 5. The accuracy level of various classifiers of machine learning and deep learning, and also other types of algorithms applied in the detection of leaf disease, are explained in Figure 5. In order to identify different types of leaf disease, several ML algorithms were used with image processing techniques.

For example, the PSO performs the lowest among all in terms of accuracy. The deep convolutional network has higher accuracy than shallow CNN network. YOLO v3 demonstrates the highest performance in comparison to other techniques.

We provide comprehensive details on various aspects of rice diseases, including preparation, distribution, exploitation, feature selection, and classification, as well as their challenges, advantages, and disadvantages. During our analysis of various literature reviews, it has come to our attention that there remains a persistent issue of limited data availability. To enhance the identification and categorization of diseases, employing any improved methodology can lead to optimal performance by effectively minimizing false classifications. By augmenting the dataset with additional images and fine-tuning the parameters of the machine learning model, it is possible to achieve even higher classification accuracy. Thorough comparisons are conducted with the existing literature.

6. Challenges

Identifying diseases in paddy fields using leaf images presents certain challenges. Addressing these challenges is essential for the development of a practical rice disease recognition system that can operate in real-time under various field conditions.

6.1. Insufficient Size and Variety of Dataset

The dataset used to train the CNN network is the main problem in many studies and articles, which leads to reduced performance reliability for the recognition of disease. A large dataset with a wide variety of photos is necessary for DL. The images were obtained in a laboratory environment on a uniform background. However, taking photographs in the field is expensive, and accurate disease recognition requires specialized agricultural skills.

6.2. Segmenting Images

Segmentation is the process of identifying the area of interest in a photograph. There are two methods for segmentation: traditional and soft computing-based. Traditional segmentation methods include K-means clustering and color thresholding, whereas soft computing-based methods include fuzzy logic, artificial neural networks, and region growth. A challenging issue in disease diagnosis is separating a leaf image from an intricate background. The segmentation of the leaf region can improve performance accuracy. Images containing many irrelevant components are frequently challenging to spot.

6.3. Image Quality

Obtaining high-quality leaf images is crucial for accurate disease identification. However, several factors can hinder image quality, such as lighting conditions, camera resolution, focus, and leaf positioning. Poor image quality can obscure important details and make it harder to discern disease symptoms.

6.4. Disease Recognition Using Visually Similar Symptoms

Certain diseases exhibit similar symptoms, making it challenging even for trained scouts to accurately differentiate them through visual inspection alone. The manifestation of disease symptoms can vary due to several factors, including geography, cultivation stage, and weather factors. To date, there has not been a study published in the literature that addresses these issues with plant disease detection.

6.5. Disease Recognition Using Real-Time Images

We found that most published works focus on the diagnosis of diseases using clinical photographs. The real-time recognition of diseases can negatively impact the performance of a model, leading to a decrease in its effectiveness. Therefore, accurate disease diagnosis in real-time field photos is a significant obstacle to overcome.

6.6. Design a Small DL Model

Hundreds of workstations, expensive GPUs, and high-performance computing systems are required for deep learning. The cost to users increases as a result. In robotics and smartphone apps, where real-time efficiency and minimal computational cost are crucial, the adoption of small convolutional neural network (CNN) models becomes highly desirable. To outperform other tactics, a sizable volume of data is required. Training is highly expensive owing to the sheer complexity of data models.

6.7. Class Imbalance

Uneven Distribution of Classes: In agricultural datasets, certain diseases may be more prevalent than others, leading to class imbalances. CNNs can be biased towards the majority class, potentially resulting in reduced accuracy for minority classes.

6.8. Environmental Variability

Impact of Environmental Conditions: Variations in environmental factors, such as lighting, humidity, and background clutter, can significantly influence the appearance of leaves. CNNs may struggle to generalize across different environmental conditions, leading to reduced robustness.

6.9. Labeling and Annotation Challenges

Annotation Consistency: Ensuring consistent and accurate labeling of training data is critical. Variability in annotators’ interpretations of disease symptoms can introduce noise into the dataset and impact the model’s performance.

7. Conclusions

In this review, we have explored and examined current shallow and deep architectures, and other ML techniques assessing their peak accuracy levels in the context of rice disease detection and crop management. The examination included an analysis of the application of realistic datasets, augmentation techniques, and diverse pre-training backbone models. While noteworthy progress has been made in this domain, an all-encompassing summarization is required to help find a more advanced and efficient approach. Automated disease detection and classification systems are needed for the exact monitoring and management of rice disease in agricultural innovations. The ML method is incorporated with image processing to detect rice leaf diseases in plants at an early stage. This paper evaluates the performances of different ML methods used for rice disease detection through image processing. With more samples, it is possible to obtain more reliable outcomes in disease detection using image processing. CNN methods extract appropriate features, including the color of the image, shape of objects, and texture of the region, for rice disease detection and classifications at an early stage. This paper summarizes different algorithms of ML and image processing and compares the performance according to the level of accuracy, dimension, and size of datasets in agricultural disease detections. New strategies can be recommended by researchers interested in this area to solve the research issue highlighted above, as well as to investigate ML and deep learning techniques that may improve the detection of rice plant diseases. Moreover, we can investigate different deep neural network architectures and fully utilize deep learning methods to enhance classification accuracy. The findings presented herein serve as a foundation for future studies, and we recommend further exploration of advanced ML and deep learning techniques to enhance disease detection capabilities.