|
Boundary-Based Rice-Leaf-Disease Classification and Severity Level Estimation for Automatic Insecticide Injection
Sayan Tepdang1, Kosin Chamnongthai1,*
Published in Applied Engineering in Agriculture 39(3): 367-379 (doi: 10.13031/aea.15257). Copyright 2023 American Society of Agricultural and Biological Engineers.
1Electronic and Telecommunication Engineering, Faculty of Engineering, King Mongkut's University of Technology Thonburi, Bangkok, Thailand.
*Correspondence: kosin.cha@kmutt.ac.th
Submitted for review on 03 July 2022 as manuscript number ITSC 15257; approved for publication as a Research Article and as part of the Artificial Intelligence Applied to Agricultural and Food Systems Collection by Community Editor Dr. Yiannis Ampatzidis of the Information Technology, Sensors, & Control Systems Community of ASABE on 29 March 2023.
Highlights
Abstract. Farmers may decide to select an appropriate insecticide for rice-leaf disease treatment in a paddy rice field based on disease class and severity level. To classify the class of rice leaf disease and estimate the severity level in a paddy rice field, several parts of the rice leaf are included in a captured image, and sometimes there exists more than one disease boundary in a part of rice leaf. This article proposes a method of rice-leaf disease classification and severity level estimation for multiple diseases on a multiple rice-leaf image. This method first finds rice-leaf candidate boundaries and identifies the rice leaf based on its feature of color, shape, and area ratio. To enlarge classification tolerance based on the coarse-to-fine concept, disease candidate boundaries are categorized into two major groups in the coarse level, and then both groups are classified into rice leaf classes in the fine level. To evaluate the performance of the proposed method, experiments were performed with 8,303 images of three rice leaf diseases including brown spot, rice blast, rice hispa and healthy rice leaf, and our proposed method achieved 99.27% which outperformed the deep learning approach by 0.43%.
Rice is the global staple food consumed by more than half of the world’s population, especially in Asia, Africa, and Latin America (Muthayya et al., 2014). It supplies approximately 20% daily calories, and 13% protein per capita. Farmers lose an estimated average of 37% of their rice crop to pests and diseases every year (Rice Knowledge Bank, 2013). Some approaches and steps exist to solve the problems of rice loss based on rice diseases. For example, practicing good cleaning of equipment and field between seasons, using clean seeds and resistant varieties, applying compatible fertilizers, encouraging natural pest enemies, properly storing grain, insecticide usage, etc. (Rice Knowledge Bank, 2013). Among those mentioned approaches most of which are considered as preventive methods, the insecticide usage approach is urgently needed to solve rice-leaf disease problems. However, there exist many types of rice leaf disease, and each disease would be killed by its specific insecticide. Therefore, insecticides must be appropriately selected based on the types of rice leaf disease, and the treatment would start from the insecticide for most rice leaf diseases in the paddy filed. Then, the insecticides are periodically changed and used for other rice-leaf diseases in the field, which may take long time depending on number of rice-leaf disease types and severity. This is expected to improve as precise agriculture to monitor rice leaf diseases and precisely treat the diseases with appropriate insecticides by real time.
Currently, farmers and experts may check the rice diseases frequently by sight. When rice diseases are found in a paddy field, farmers urgently select suitable insecticide for treatment to protect rice growth yield. If there are multiple rice diseases in a paddy field, farmers might select a single type of insecticide, which is matched with only one or a few types of rice diseases, to treat the whole paddy field. Although most of rice plants in the paddy field are treated due to the usage of a type of insecticide, some may remain to damage rice plants and decrease rice growth yield. Therefore, it is crucial to precisely classify rice disease types in every leaf in order to set an appropriate plan to protect rice plants from all diseases. If we can precisely detect rice diseases, paddy field zoning for suitable insecticide treatment can be realized in order to deal with all types of rice diseases. Besides rice-leaf disease classification, another important factor to decide on insecticide injection is severity level, which is different according to the rice-leaf disease type (Chettanachit et al., 2007; Plantwise Knowledge Bank., 2013; Rice Knowledge Bank., 2013). It is also really needed to develop a function of rice-leave-disease severity-level estimation for the development of automatic insecticide injection. Recently, some automatic systems and robots are recently developed to inject insecticide into rice leaves (Ganesh et al., 2020). These automatic insecticide injection systems and robots require intelligent functions such as rice disease classification, severity level estimation, and so on to be improved at an autonomous level.
On the other hand, precision agriculture recently emerged and efficiently controlled modern food production including rice planting (Ullah et al., 2017; Mehta and Teeyasuksaet, 2018; Shafi et al., 2019). Based on the concept of precision agriculture, rice diseases should be detected and precisely treated. In fact, these multiple rice diseases sometimes reside in the same paddy field and in the single rice leaves, and absolutely require different treatments (Chettanachit et al., 2007; Tollenaere, et al., 2017; Cunnac, 2021) against each disease. Therefore, the detection and classification of rice leaf disease boundaries in an image became our research problem. Then, these detected and classified rice leaf disease boundaries are useful for disease zoning and severity level estimation.
Existing rice-leaf-diseases classification methods can be categorized into two groups which are pattern recognition and machine learning-based approaches. In the first group of pattern recognition (A.Hussein et al, 2018; Bakar et al., 2018), all patterns are initially supervised and will be classified. These methods work well in the trained patterns without any learning functions. On the other hand, machine learning emerged later and utilized in many methods. Disease detection and classification methods based on machine learning (Prajapati et al., 2017; Al-Amin et al., 2019; Hasan et al., 2019; Shrivastava et al., 2019; Feng et al., 2020; Ghosal and Sarkar, 2020; Matin et al., 2020; Pothen and Pai, 2020; Sethy et al., 2020; Ramesh and Vydeki, 2020; Mekha and Teeyasuksaet, 2021; Latif et al., 2022) have been developed for plant leaves including rice, cucumber, maize, wheat, etc. These methods could classify disease in an image. On the other hand, a method using deep learning-based faster R-CNN framework (Islam et al., 2021), which is classified at an automatic feature extraction approach, could excellently detect and classify multiple rice leaf diseases in an image.
However, the automatic feature-extraction-based method basically requires big data for training, and rice leaf diseases are varied based on area, climate, season, and so on which is difficult to find big data for training. Hence, traditional feature design methods become an effective choice. Normally, a rice-leaf disease boundary is possibly depicted in any shapes. It is needed to develop an algorithm to precisely segment boundaries according to their shapes and classify those boundaries into rice leaf disease. The precise area of rice-leaf disease boundary can be used to estimate severity level which is an important factor to decide on insecticide injection. Therefore, this article develops a classification algorithm of multiple rice-leaf disease on multiple rice leaves in an image for automatic insecticide injection.
The remainder of this article is organized as follows: Section 1 introduces motivation, background, existing related works, and research problem; Section 2 provides details of analysis of rice diseases and concept; Section 3 describes proposed rice-leaf-disease classification method; Section 4 explains the implementation of rice-leaf-disease classification; Section 5 reports details of experiment implementation, data set design, empirical results evaluation, and discussion; and Section 6 concludes the entire work.
Analysis of Rice-Leaf Diseases and the Concept
 | ||||||
| (a) | (b) | (c) | ||||
| Figure 1. Images of rice leaves: (a) non-disease (b) single disease (c) multiple disease. | ||||||
Normally, rice paddy trees are planted in a field, and rice leaves can be seen randomly (fig. 1a). When rice leaf is attacked by diseases, the boundaries of rice leaf diseases can be seen on rice leaves. A rice leaf can be attacked by a single disease and multiple diseases, respectively (fig. 1b and 1c). Since insecticides for effective rice-leaf disease treatment should be selected properly based on rice leaf diseases (table 1), rice leaf diseases are required to be precisely detected based on boundaries on rice leaves.
Suppose rice leaf diseases are extracted based on boundaries, there exit approximately 10 rice leaf diseases (Rice Knowledge Bank, 2013). For example, the top 5 rice leaf diseases are rice blast, narrow brown spot, bacterial blight, brown spot, and bacterial leaf streak (fig. 2). If we define Z as a set of rice leaf diseases, Zi ; i = 1, 2, 3, …, n represents all (n)types of rice leaf diseases.
When we extract some features such as shape factor and circle on some samples of those rice leaf diseases, they are distributed in a feature coordinate space (fig. 3). If e is defined as a coordinate in the feature coordinate space, ej ; j = 1, 2, 3, …, m represents all coordinates of a set of a rice leaf disease and ej?Zi. A set of a rice leaf disease can be expressed as:
(1)
where
i and j = number of rice-leaf disease types and number of
samples in a rice leaf disease, respectively.
Then, the centroid (
) of all samples in a set of rice leaf disease can be expressed as:
(2)
where
m = represents number of coordinates of a rice leaf disease.
Statistically, some samples considered as insignificant ones so that these should be deleted. In this article, 6s (Wikipedia, 2022), which is popularly accepted as a powerful technique and tool, is proposed to filter the irregular samples. The standard deviation can be expressed by:
(3)
Then, irregular samples are rejected by:
(4)
In figure 3, rice-leaf disease types can be drawn boundaries from arbitrary shape, which are difficult to express in a mathematic model. Hence, this article simply approximates these boundary shapes into a circle covering all samples in a group.
To cover all samples in a set of rice leaf diseases, the radius (ri) of the circle becomes the distance from the centroid to the farthest sample.
(5)
where
n and m = rice-leaf disease type (e.g., brown spot, rice blast,
bacterial leaf streak, bacterial blight, narrow
brown spot, etc.) and number of samples in a rice-
leaf disease group, respectively.
On the other hand, a distance (dL) between a couple of rice leaf diseases can be obtained by the distance between circles representing rice-leaf disease sets, as shown in figure 3. Therefore, the distance can be calculated as:
(6)
where
? and L = the distance between centroids of a couple of
circles and a couple of circles (e.g., AB, AC, AD,
AE, BC, BD, BE, CD, CE, etc.), respectively.
Finally, these distances (dL ) among sets of rice
leaf disease are considered to classify into
multiple groups.
| Table 1. Rice leaf diseases and insecticides. | ||||
| Diseases | Image Example | Diseases Boundary | Severity Level as Condition for Insecticide Treatment | Insecticide |
| Brown spot |  |  | 6%-10% | Mancozeb, Carbendazim |
| Rice blast |  |  | 4%-5% | Tricyclazone, Kasugamycin |
| Bacterial leaf streak |  |  | 1%-3% | Nitrogen |
| Bacterial blight |  |  | 6%-10% | Bacbicure |
| Narrow brown spot |  |  | less than 1% | Carbendazim |
 |
| Figure 2. Rice leaf diseases. |
To enhance tolerance among groups in feature domain in order to increase classification accuracy, rice-leaf disease classes must be considered to merge into the same class in the case of short distance. This article merges all classes in term of pair. Pairs (P) of classes (Z) are established in:
 |
| Figure 3. Distances of rice leaf disease groups. |
(7)
where
m = represents number of classes.
Pairs of rice-leaf-disease classes that are partially overlapped (e.g., Groups A and B) and the ones with short distances (e.g., Groups C, D, and E) can be considered to merge in the same group (Gj), as shown in equations 8 and 9, respectively.
(8)
(9)
where
j and th = number of new merged groups and a threshold for
group merging, respectively.
Therefore, users may collect some samples of rice leaf diseases to perform training according to mentioned equations 1-9 in advance. As an example, circles (A-E) representing boundaries of rice-leaf disease types are established in the feature coordinates, and merged groups surrounded by green circles are determined, as shown in figure 4 in the training state. These rice leaf diseases in merged groups are used to train classifiers separately, and the trained classifiers are used for classification in the testing state.
In the testing state, as coarse step, candidate rice-leaf-disease boundaries (eij), which are located in the feature coordinates are tentatively classified into one of rice-leaf disease types (e.g. A-E) by covering by rice-leaf-disease type circle or the shortest distance (
) between the boundary (eij) and the centroid (Ci) in the feature coordinates as expressed by:
(10)
where
= distance from a coordinate (eij) to the centroid (Ci)
of a rice leaf disease.
As fine step, tentatively classified rice-leaf diseases are separated for different classifiers according to merged groups.
 |
| Figure 4. Rice-leaf disease groups in the feature coordinate. |
Proposed Rice-Leaf-Disease Classification Method
Based on the mentioned analysis results and concept, our proposed method of rice-leaf-disease classification is established and explained in this section. The section is structured as follows: the first two sub-sections introduce overall proposed method of rice-leaf disease classification and training state, respectively; the third sub-section mentions leaf boundary segmentation, leaf confirmation, and candidate rice-leaf-disease boundary detection; the fourth and fifth sub-sections explain feature-based disease classification and group merging, respectively; and the sixth and seventh sub-sections discuss deep learning classification and estimation of severity of intensive effects, respectively.
Overall Proposed Method of Rice-Leaf Disease Classification
Overall proposed method of rice-leaf disease classification is depicted in term of flowchart (fig. 5). It is divided into training and testing states located on the left and right sides, respectively. The system starts by inputting an image of rice leaves in the rice paddy field. In the image, candidate leaf boundaries are segmented by a segmentation tool, leaf boundaries are confirmed based on color information, and candidate rice-leaf-disease boundaries are detected in the rice leaf boundaries in which the details are explained in algorithms 1, 2, and 3, respectively. These mentioned processes are common between training and testing states.
In training state, the candidate rice-leaf-disease boundaries are classified in the coarse level into tentative rice-leaf disease groups based on equation 10 in which details are explained in algorithm 4. Those tentatively classified groups are merged based on algorithm 5, and all merged groups are trained on specific deep learning classifiers at the fine level.
The details of the training state are mentioned in the next sub-section.
On the other hand, the testing state written on the right side in figure 5 is performed as follows. Based on equation 10, the candidate rice-leaf-disease boundaries are classified in the coarse level into merged rice-leaf disease groups in the process of feature-based disease classification in which details are explained in algorithm 6. Then, those merged groups are classified into rice-leaf-disease classes using specific deep learning classifiers, which have been trained in the training state. The details of deep learning classification are discussed in the second last sub-section. Finally, the severity of intensive effects is simply estimated based on ratio of areas between rice leaf and rice leaf diseases, and details are explained in algorithm 7.
 |
| Figure 5. Flowchart of overall rice-leaf disease classification. |
| Algorithm 1. Rice-leaf boundary segmentation. |
1: BEGIN |
2:COMPUTE: Conversion of a color image into a binary image |
3: COMPUTE: Segmentation of a color image into green, magenta, yellow and red channel image. |
4: COMPUTE: Sum of product of binary image, and green, magenta, yellow and red images |
5: END |
Algorithm 2. Rice leaf confirmation. |
1: BEGIN |
2: INPUT: A segmented image of candidate rice leaf parts |
3: COMPUTE: Edge detection |
4: COMPUTE: Texture detection |
5: COMPUTE: Classification using edge, texture, and color |
6: END |
Algorithm 3. Disease-candidate-boundary detection. |
1: BEGIN |
2: INPUT: Rice-leaf boundary images including rice leaves and rice leaf diseases |
3: COMPUTE: Segmentation of rice-leaf disease boundaries in the rice leaf boundaries using K-mean |
4: COMPUTE: Classification of rice-leaf disease candidates in the segmented boundaries |
5: END |
Algorithm 4. Determination of rice-leaf-disease type boundaries on feature-based coordinates. |
1: BEGIN |
2: SET Coordinates of rice-leaf disease samples |
3: COMPUTE: Centroids of all rice-leaf disease classes |
4: COMPUTE: Standard deviation of all rice-leaf disease classes |
5: COMPUTE: Rejection of irregular samples based on 6s |
6: COMPUTE: Radius of rice-leaf disease boundaries based on the sample with maximum distance in the class |
7:END |
Training State
In advance, rice-leaf disease boundaries on some samples must be manually classified by rice-leaf disease experts, and those classified images of rice-leaf diseases would be handled by users to train classifiers for obtaining weights and parameters for rice-leaf disease classification in testing state. Naturally, rice leaf diseases have similar features, sometimes it is difficult to classify rice-leaf disease classes due to small tolerances among classes. Our basic concept recommends enhancing the tolerances by applying coarse-to-fine concept. Comprehensively, the processes in the training state are depicted in the left side of figure 5. In the process of determination of rice-leaf-disease type boundaries on feature-based coordinates, boundaries of rice-leaf-disease groups, which are assumed to be circle shape, are determined by circle centroid and radius according to equations 2 and 5. Then, rice-leaf-disease groups are considered to be merged in the coarse level based on distances among rice-leaf-disease groups on the feature-based coordinates according to equations 8 and 9. The details of group merging are described in algorithm 5. Finally, all merged groups are fed to different deep learning classifiers for training and obtaining weights and parameters.
| Algorithm 5. Group merging. |
1: BEGIN |
2: SET di: distance between a pair of classes, Si : slope of distances, N : number of classes |
3: COMPUTE: Reordering of di from minima to maxima |
4: FOR i =1 to N -1 do |
5: CALCULATE: slope (Si ) |
6: END FOR |
7: CALCULATE: Max (Si ) as border between merged and non-merged groups |
8: CALCULATE: Merging rice-leaf disease classes in the merged group |
9: END |
Algorithm 6. Feature-based rice-leaf-disease classification |
1: BEGIN |
2: SET N : total number of testing samples, M : total number of rice-leaf disease classes, |
3: FOR i =1 to N do |
4: FOR j =1 to M do |
5: CALCULATE: distance ( |
6: END FOR |
7: FOR j =1 to M do |
8: IF |
9: ELSE IF ( |
10: END FOR |
11: END FOR |
12: END |
Algorithm 7. Estimation of severity intensive effects. |
1: BEGIN |
2: INPUT: An image of rice leaf |
3: INPUT: An image of disease boundaries in the rice leaf |
4: COMPUTE: Total area of rice leaf |
5: COMPUTE: Area of infected region |
6 COMPUTE: Ratio of area of infected region and total area of rice leaf |
7: END |
: distance between a point to class centroid
)
then 
= MIN 
) then
Leaf Boundary Detection, Leaf Confirmation, and Candidate Rice-Leaf-Disease Boundary Detection
In practice, a digital image (f(x,y)) captured by a camera installed in a rice paddy field complicatedly includes rice leaves with rice-leaf disease boundaries and background (fig. 6a). The system converts the image into a binary image (b1(x,y)) based on a thresholding value (th1) which has been determined in advance. An example is shown in figure 6b, and the binary image can be simply expressed as:
(11)
The binary image (b1(x,y)) including background (0) and foreground (1) boundaries would be later used as a window for filtering rice leaf parts in the input image (f(x,y)). Simultaneously, the rice leaf and rice-leaf disease parts are segmented on the input digital color image (f(x,y)) into rice leaf and rice-leaf disease images using green (g(x,y)) channel, and magenta (m(x,y)), yellow (w(x,y)), and red (r(x,y)) channels, respectively (fig. 6c). These four images, which are assumed to include rice leaf and rice-leaf disease boundaries, are then multiplied by the binary window image to obtain three images including rice leaf and rice leaf disease parts with background. Subsequently, these three images are operated sum merged into an image (s(x,y)) of rice leaf parts with rice leaf diseases as:
(12)
where
s(x,y) = a segmented image.
The procedures for rice leaf boundary detection, rice leaf confirmation, and candidate rice-leaf-disease boundary detection are introduced in Algorithms 1-3, respectively. In algorithm 1, an input image is binarized by an appropriate threshold value into a binary image representing rice leaf position (fig. 6b). The binary image is later used as a filter for extracting rice leaf parts. Simultaneously, the input color image is segmented using color information into green channel, magenta, yellow and red channels (fig. 6c). The images extracted by green channel, and magenta, yellow, and red channels are considered to represent rice leaf and rice leaf diseases. Finally, sum of product between the binary image and other images is performed to find a segmented rice leaf image according to equation 11. An example of the segmented rice leaf image is shown as rice leaf candidate in figure 6d.
In algorithm 2, the rice-leaf part candidates in the segmented image obtained by algorithm 1 would be confirmed as rice leaf. In this article, edge, texture, and color information, as shown in figures 7a-c, respectively, is used as features to train and test in a deep learning classifier. The segmented candidate rice-leaf image is used to extract edge, texture, and color information to feed to the classifier for rice leaf classification as rice leaf confirmation.
In algorithm 3, rice leaf parts confirmed by algorithm 2 are segmented to find rice-leaf disease boundaries. First, the confirmed rice leaf images are input and used to segment into candidate rice-leaf-disease boundaries using a classifier. To establish a classifier, users need to analyze images of rice leaf and rice leaf diseases to find effective features in prior. In the case of rice leaf diseases, color information is obviously observed to be able to differentiate rice leaf disease from rice leaf. As an example, green and yellow colors, which are considered as major components in rice leaf and rice-leaf disease boundaries, as shown in figure 8, are analytically categorized into rice leaf and disease pixels so that these features should be selected and used to label on disease pixels.
In classifier selection, users may carefully consider selecting an appropriate classification tool out of existing ones. In this article, K-means which is regarded as simple and friendly classification tool is recommended to use as classifier for the boundary candidate of rice leaf disease. Some samples are needed to train the selected classifier such as K-means, and so on for determination of thresholding parameter in advance. The trained classifier would be applied for segmentation of candidate rice-leaf-disease boundaries on the confirmed rice leaf boundaries into rice-leaf disease classes in the testing state. As an example, candidate rice-leaf-disease boundaries are classified using K-means into rice-leaf disease and rice leaf boundaries based on yellow and green colors (fig. 8). Finally, the rice-leaf disease boundaries (fig. 9) are classified into rice-leaf disease classes (e.g., brown spot, rice blast, bacterial leaf streak, bacterial blight, narrow brown spot, etc.) by deep learning classifier.
 |  |  |  |
| (a) | (b) | (c) | (d) |
| Figure 6. Output examples of processes in the algorithm 1. (a) Input image, (b) binary image, (c) green, magenta, yellow, and red channels, and (d) segmented image. | |||
 |  |  |
| (a) | (b) | (c) |
| Figure 7. Extracted (a) edge, (b) texture, and (c) color. | ||
 |
| Figure 8. Yellow and green channels as features of rice leaf and disease boundaries. |
 |
| Figure 9. Detection of rice-leaf disease boundaries. |
Determination of Rice-Leaf-Disease Boundaries on Feature-Based Coordinates
The process of determination of rice-leaf-disease boundaries on feature-based coordinates aims to determine boundaries of rice-leaf disease classes on feature-based coordinates in the training state. The determined boundaries would be used to classify rice leaf disease boundaries into rice-leaf disease classes in the testing state. The procedure follows algorithm 4. First, samples of rice leaf diseases classified by experts are collected and analyzed to determine coordinates in the feature-based coordinates. In this article, shape factor and circle are selected as feature for rice-leaf-disease class classification due to their obvious differences among rice-leaf-disease classes. These features are plotted in the feature coordinates, as shown in figures 3 and 4. Then, centroids of all rice-leaf disease classes are obtained from their samples, and the centroids would be used as representative of rice leaf disease classes in the feature coordinates. Statistically, some samples, which are apart from the centroids should be excluded as irregular ones so that standard deviation is calculated according to equation 3, and some samples considered as irregular ones are rejected based on 6s according to equation 4. Simply, the boundaries of rice-leaf disease classes are assumed as circle shape in this article so that the samples with maximum distances from the centroids obtained by equation 5 are used as radius to draw the circle boundaries of the rice-leaf disease classes.
Group Merging
Group merging is described in algorithm 5. It starts from reordering all pairs (P) of the rice-leaf disease classes (Z) by their distances from minima to maxima in the step 3. Then, the system starts a loop for computing slopes of consecutive reordered distances of pairs, and the biggest slope is selected from all slopes (si) as border between merged and non-merged groups. This means the border extracts a group of pairs with short distances located in the left side as merged classes. Finally, those rice-leaf disease classes including in the pairs in the merged classes would be merged in the same classes for the coarse classification step.
An example based on the case in figure 3 is considered for group merging and the results are shown in figure 10. In this case, all possible pairs are reordered from minima to maxima, their slopes (si) between consecutive pairs are obtained, and s3 is the highest slope and selected as the border. As results, AB, CD, and DE pairs are selected to be merged in the same groups, and A and B classes, and C and D and E classes become two groups for the coarse classification step in the training state. Those two groups are surrounded by green lines in figure 4.
 |
| Figure 10. Distances of disease-class pairs reordered from minima to maxima. Need to renumber figures |
Feature-Based Rice-Leaf-Disease Classification
In the process of feature-based rice-leaf-disease classification in testing state, distances (
) from a coordinate (ei) of a testing sample to all rice-leaf disease centroids (cj) in the feature-based coordinates are measured, those measured distances are evaluated based on equation 10. If the distance from the coordinate (ei) of a testing sample to a rice-leaf disease centroid is shorter than the radius of rice-leaf-disease class boundary, the testing sample is included in the rice-leaf disease class. Otherwise, the testing sample would be included in the rice-leaf disease class with the minimum distance (algorithm 6).
Deep Learning Classification of Disease
Recently, many kinds of deep learning have been developed, and they can deal well with many applications. Users are recommended to survey and find an appropriate one for applying in the system in advance. In this article, Resnet50, which was recommended as effective tool (Jung, 2018; Gurevin et al., 2021) was used as classifier in the system. Since the structure comprises of input, hidden, and output layers, number of nodes in those layers are surveyed and performed pretest in order to set the whole network structure, and appropriate parameters for running classification are found in the training state. A set of parameters for the Resnet50 is introduced as an example in table 2. In the first item of image size, users are free to set image size, but need to consider memory in the hardware and processing time (Clarke, 2019). In the item of batch size which is working space, it should be considered based on conditions of ability of GPU and memory (Atrio and Popescu-Belis, 2022). In dropout, which helps protecting overfitting, it is recommended as 0.5 (Jabir and Noureddine, 2021). In the item of epochs which is learning times, it is recommended as 50 times (Gurevin et al., 2021). In optimizer which is learning rate, Adam is recommended as efficient one (Loris et al., 2021). In the activation function, the parameter can be selected from existing ones such as softmax, relu, sigmoid, tanh, and so on. Normally, sigmoid is simple and popularly used, and it can be used in this article. However, softmax which looks close to sigmoid is selected and used in this article due to its effectiveness. Finally, output class is determined by the number of classes, which are brown spot and rice blast group or bacterial leaf streak, bacterial blight, and narrow brown spot group.
| Table 2. Parameters set in a deep learning network for brown spot, rice blast, bacterial leaf streak, bacterial blight, and narrow brown spot classes. | ||
| Network Layers | Parameter Options | Selection |
| Input | Image_size | 224*224 |
| Batch_size | 8 | |
| Dropout | 0.5 | |
| Hidden layer | Number_epochs | 50 |
| Optimizer | Adam | |
| Activation Function | Softmax | |
| Output layer model | Output | 2 Class or 3 Class |
Estimation of Severity Level
According to the study results about rice disease (Chettanachit et al., 2007; Plantwise Knowledge Bank., 2013; Rice Knowledge Bank., 2013) when rice diseases occur on rice leaves, lesions may appear on the leaves. This basically affects the rice plants from maximum tillering to panicle initiation, and rice plants unfortunately may experience grain yield losses. Although rice disease treatment using an insecticide is not recommended due to consumer health reasons, it must be performed when reaching unavoidable severity levels.
The rice-disease severity levels or severity of intensive effects (?) are basically determined based on the area ratio between disease boundaries or area of infected region and total area of leaf (Rice Knowledge Bank, 2013), and it can be simply expressed as:
(13)
Hence, rice diseases should be detected, classified, and a proper treatment should be performed in order to prevent immense damages. In the case of unavoidable severity levels as shown by conditions in the 4th column in table 1, an appropriate insecticide for each rice leaf disease (e.g., brown spot, rice blast, bacterial leaf streak, bacterial blight, narrow brown spot, etc.), as shown in the 5th column is recommended to use for the disease treatment based on the standard evaluation system (SES) of the International Rice Research Institute (2002). To obtain rice-disease severity level, procedure for obtaining rice-leaf severity levels simply follows the algorithm 7, and examples of output obtained in algorithm 7 are shown in figure 11.
 |  |  |
| (a) | (b) | (c) |
| Figure 11. Examples in processes in rice-disease severity evaluation. (a) Image of rice leaf, (b) image of disease boundaries in the rice leaf, and (c) area of the disease boundaries. | ||
Experimental Results
To evaluate performance of the proposed method, images of multiple rice leaves and rice leaf diseases were collected by hardware system based on the specification, as shown in table 3. As image samples for experiments, one disease (brown spot or rice blast) in single leaf images, one disease (bacterial leaf streak or bacterial blight) in single leaf images, multiple diseases in single leaf images, and multiple diseases with multiple leaves in images have been captured, as shown in table 3 (Dataset, 2021), respectively. The single leaf images were trained and tested for comparison with conventional methods, and the second and third groups of multiple rice leaves with multiple diseases were classified experimentally using existing tools based on the proposed method.
Results of the detected rice-leaf disease boundaries of the proposed method are shown in table 4. To compare with the conventional methods (Islam et al., 2021), single-leaf and multiple-disease images were experimentally performed in healthy rice leaf and three rice leaf diseases, which were brown spot, rice blast, and rice hispa in the 50:50 rate of training and testing, respectively. The classification accuracy is shown in table 4. To evaluate performance of the proposed method from various training and testing ratio, experimental results are performed in 90:10, 80:20, 70:30, 60:40, and 50:50 using accuracy (ACC) as standard deviation (SD), as shown in table 5.
| Table 3. Experiments specifications and image data. | |||
| Data/Devices/ Software/ | Specifications | Image Example | |
| Data | Rice leaf disease: 4,000 images Disease boundaries: 5,000 images Multiple diseases: 1,000 images Multiple rice leaf: 500 images |  |  |
| Single disease and single leaf images (brown spot or rice blast classes) | Single disease and single leaf images (bacterial leaf streak, bacterial blight, and narrow brown spot classes | ||
 |  | ||
| Multiple diseases and single leaf images | Multiple diseases and multiple rice leaves | ||
| Computer system | HP PROBOOK 440G5 CPU: Intel Core i5-82500U Memory size: 16 GB Hard disk drive: 1 TB |  | |
| Rice disease boundaries in classes. brown spot or rice blast and bacterial leaf streak classes | |||
| Software | Python, MATLAB, Rapid Miner Studio |  | |
| Rice disease boundaries in classes bacterial blight and narrow brown spot classes | |||
| Table 4. Comparison of disease classification. | |||
| Rice Diseases | Conventional Method (Islam et al., 2021) | Proposed Method | Standard Deviation |
| Brown spot | 98.09 | 98.55 | 0.03 |
| Rice blast | 98.85 | 98.89 | 0.01 |
| Rice hispa | 99.17 | 99.70 | 0.01 |
| Healthy rice leaf | 99.25 | 99.97 | 0.04 |
| Avg. | 98.84 | 99.27 | 0.02 |
| Table 5. Classification accuracy in the ratio of training and testing according to diseases. | ||||||||
| Training and Testing | Diseases | ProposedCoarse-to-Fine | Training and Testing | Diseases | Proposed Coarse-to-Fine | |||
| Accuracy | Standard Deviation | Accuracy | Standard Deviation | |||||
| 90:10 | Brown spot | 97.90 | 0.02 | 70:30 | Bacterial blight | 96.13 | 0.05 | |
| Rice blast | 98.30 | 0.02 | (Con) | Narrow brown spot | 94.30 | 0.08 | ||
| Bacterial leaf streak | 95.70 | 0.03 | AVG | 95.60 | 0.06 | |||
| Bacterial blight | 96.20 | 0.03 | 60:40 | Brown spot | 97.52 | 0.05 | ||
| Narrow brown spot | 96.00 | 0.03 | Rice blast | 96.79 | 0.05 | |||
| AVG | 96.82 | 0.03 | Bacterial leaf streak | 91.37 | 0.9 | |||
| 80:20 | Brown spot | 98.35 | 0.02 | Bacterial blight | 95.47 | 0.8 | ||
| Rice blast | 97.75 | 0.03 | Narrow brown spot | 93.92 | 0.05 | |||
| Bacterial leaf streak | 95.15 | 0.04 | AVG | 95.01 | 0.07 | |||
| Bacterial blight | 96.70 | 0.03 | 50:50 | Brown spot | 97.34 | 0.05 | ||
| Narrow brown spot | 95.35 | 0.04 | Rice blast | 95.50 | 0.07 | |||
| AVG | 96.66 | 0.03 | Bacterial leaf streak | 90.02 | 0.11 | |||
| 70:30 | Brown spot | 97.93 | 0.04 | Bacterial blight | 93.64 | 0.12 | ||
| Rice blast | 97.32 | 0.04 | Narrow brown spot | 93.72 | 0.06 | |||
| Bacterial leaf streak | 92.33 | 0.09 | AVG | 94.04 | 0.08 | |||
Furthermore, experiments have also been performed in different eight countries with 100 rice leaf images per country (Datacountry, 2021), and in different rice types, locations, and seasons with 1,000 images per disease, the evaluation results are shown in tables 6 and 7, respectively. The evaluation results of severity level estimation compared with conventional method (Tosawadi et al., 2021) are shown in table 8. Columns 2-5 and 6-9 in the table 8 show severity levels (S) in four classes (0 < S « 4%, 4% < S « 6%, 6% < S « 11%, and S > 11%), respectively, while rows show five rice leaf diseases. In severity level estimation, 80 rice leaves per severity level per disease, totally 1,600 leaves are used, and results show our proposed estimation based on area outperforms the conventional method using deep learning 4.07% approximately.
| Table 6. Classification accuracy among countries. | |||||||
| Country | Diseases | Avg. | Standard Deviation | ||||
| Brown Spot | Rice Blast | Bacterial Leaf Streak | Bacterial Blight | Narrow Brown Spot | |||
| India | 97.70 | 98.65 | 93.25 | 93.70 | 93.50 | 95.36 | 0.02 |
| Thailand | 98.70 | 97.65 | 96.20 | 93.70 | 94.00 | 96.05 | 0.02 |
| USA | 96.20 | 98.65 | 97.70 | 97.70 | 94.00 | 96.85 | 0.01 |
| Vietnam | 96.20 | 97.65 | 94.90 | 97.70 | 97.75 | 96.84 | 0.01 |
| China | 98.70 | 98.65 | 96.15 | 96.20 | 97.75 | 97.49 | 0.02 |
| Myanmar | 96.20 | 98.65 | 94.90 | 97.70 | 92.75 | 96.04 | 0.02 |
| Brazil | 98.70 | 98.65 | 93.65 | 98.70 | 94.00 | 96.74 | 0.02 |
| Cambodia | 96.20 | 98.65 | 93.65 | 96.20 | 93.50 | 95.64 | 0.02 |
| Avg. | 97.32 | 98.40 | 95.05 | 96.50 | 94.65 | 96.38 | 0.02 |
| Table 7. Classification accuracy in seasons and rice variety. | ||||||
| Disease | Varieties | Season | Place | Time | Avg. | Standard Deviation |
| Brown spot | RD41 RD49 Khao Dawk Mali 105 (two disease) | Rainy Rainy Winter | Chon Buri Chon Buri Chon Buri | Aug. 2019 Aug. 2019 Dec. 2019 | 98.60 99.20 97.50 | 0.05 0.04 0.06 |
| Rice blast | RD51 RD47 | Rainy Summer | Bangkok Chachoengsao | Jul. 2019 Feb. 2020 | 98.40 99.30 | 0.06 0.04 |
| Bacterial leaf streak | Phitsanulok 2 | Rainy | Phichit | Sep. 2019 | 95.60 | 0.03 |
| Bacterial blight | RD49 Khao Dawk Mali 105 (two disease) RD47 (two disease) | Rainy Winter Winter | Kamphaengphet Chon Buri Chon Buri | Sep. 2019 Dec. 2019 Jan. 2020 | 98.20 96.20 95.70 | 0.05 0.07 0.07 |
| Narrow brown spot | RD31 RD47 (two disease) | Rainy Winter | Chon Buri Chon Buri | Aug. 2019 Jan. 2020 | 98.50 97.20 | 0.07 0.08 |
| Avg. | 97.67 | 0.05 | ||||
Discussion
In this article, rice leaf diseases have been classified in a paddy field for appropriate selection of insecticide, and a rice-leaf-disease classification method has been proposed for rice leaf images including multiple rice leaf diseases and multiple rice leaves. The proposed method first performs boundary detection in a rice leaf image, detects rice leaves, and determines rice-leaf disease boundaries in the detected rice leaf boundaries in the preprocessing. Then, rice leaf diseases are classified based on the detected rice-leaf disease boundaries so that the proposed method can deal with images including multiple leaves in an image and multiple diseases in a leaf. Furthermore, classification in the method is performed based on coarse-to-fine concept which is similar to human ability and helps classifying precisely.
As shown in table 4, experimental results done by the conventional method (Islam et al., 2021) using deep learning reveal 98.84% approximately, while the results of our proposed method using random forest and deep learning show 99.27%. This means our proposed method comprehensively can improve accuracy compared with the conventional method in the same conditions. As checked the accuracy of classification in the coarse step, the accuracy is found as high as 99.48%, which is close to 100%. The high classification accuracy in the coarse step is considered to enlarge tolerance among classes and lead to achieve approximately 94% to 96% accuracy (table 5). From the recognition rates of rice-leave disease classes, it is obvious that brown spot and rice blast classes gain much better accuracy than bacterial leaf streak, bacterial blight, and narrow brown spot classes, since features in the latter group of bacterial leaf streak, bacterial blight, and narrow brown spot such as bright brown color and straight-line shape look similar. Other features in this group should be analyzed and added to the classifier as future works. To confirm the performance of the proposed method in different planting locations, seasons, and rice species, experimental results (tables 6 and7) show that rice leaf diseases can be classified by our proposed method by as high as 95% to 99% accuracy approximately.
| Table 8. Severity level estimation. | ||||||||||
| Severity Level (S) | ||||||||||
| Conventional Method (Tosawadi et al., 2021) | Proposed Method | |||||||||
| Diseases | 0 < S « 4% | 4% < S « 6% | 6% < S « 11% | S > 11% | 0 < S « 4% | 4% < S « 6% | 6% < S « 11% | S > 11% | ||
| Brown spot | 90.00 | 92.50 | 92.50 | 93.80 | 95.00 | 93.80 | 95.00 | 96.20 | ||
| Rice blast | 93.80 | 92.50 | 93.80 | 95.00 | 97.50 | 97.50 | 96.20 | 98.80 | ||
| Bacterial leaf streak | 86.20 | 88.80 | 86.60 | 90.00 | 90.00 | 92.50 | 93.80 | 95.00 | ||
| Bacterial blight | 93.80 | 92.50 | 90.00 | 93.80 | 96.20 | 96.20 | 96.20 | 97.50 | ||
| Narrow brown spot | 88.80 | 86.20 | 88.80 | 92.50 | 92.20 | 92.50 | 95.00 | 96.20 | ||
| Table 9. Confusion matrix. | ||||||
| Ground Truth | ||||||
| Diseases | Brown Spot | Rice Blast | Bacterial Leaf Streak | Bacterial Blight | Narrow Brown Spot | |
Classification Results | Brown spot | 989 | 11 | 0 | 0 | 0 |
| Rice blast | 7 | 993 | 0 | 0 | 0 | |
| Bacterial leaf streak | 0 | 0 | 967 | 15 | 18 | |
| Bacterial blight | 0 | 0 | 18 | 972 | 10 | |
| Narrow brown spot | 0 | 0 | 13 | 17 | 970 | |
| Table 10. Classification error. | ||||||||||||
| Classification Error | Rice Blast ? Brown Spot | Brown Spot ? Rice Blast | Bacterial Blight ? Bacterial Leaf Streak | Narrow Brown Spot ? Bacterial Leaf Streak | ||||||||
| Shape | Histogram | Shape | Histogram | Shape | Histogram | Shape | Histogram | |||||
| Ground Truth |  |  |  |  |  |  |  |  | ||||
| Classification Results |  |  |  |  |  |  |  |  | ||||
| Classification Error | Bacterial Leaf Streak ? Bacterial Blight | Narrow Brown Spot ? Bacterial Blight | Bacterial Leaf Streak ? Narrow Brown Spot | Bacterial Blight ? Narrow Brown Spot | ||||||||
| Shape | Histogram | Shape | Histogram | Shape | Histogram | Shape | Histogram | |||||
| Ground Truth |  |  |  |  |  |  |  |  | ||||
| Classification Results |  |  |  |  |  |  |  |  | ||||
Moreover, classification errors are categorized in a confusion matrix (table 9) in which ground truth and classification results are allocated in column and row directions, respectively. Most of classification results are corrected as shown in diagonal direction elements, but other elements surrounded by circles show number of errors in eight groups (rice blast ? brown spot, brown spot ? rice blast, bacterial blight ? bacterial leaf streak, narrow brown spot ? bacterial leaf streak, bacterial leaf streak ? bacterial blight, narrow brown spot ? bacterial blight, bacterial leaf streak ? narrow brown spot, and bacterial blight ? narrow brown spot). All classification errors in the eight groups are analyzed to find causes, and representatives of those groups are shown in table 10. As observed in table 10, although histograms look close to the ones of the ground truth, their shapes are closer to others, as shown by red lines. Analytically, the ground truth results, which are originally expected to be the classification results, failed to get the top score, but most of them can be kept as the second one in classification results. This supports the thought that if some features are added in the future, the classifier is considered to be improved and may obtain better accuracy. Furthermore, severity level of diseases was estimated based on equation 13, and the results indicated an accuracy of 95.17% (table 8). Since images of rice leaf in experiments are captured randomly, some parts of rice leaf images are occluded. The area of whole rice leaves and all their disease boundaries are inevitably regarded to be missed. If areas of both rice leaf and rice-leaf disease boundaries are decreased in the same ratio, the ratio estimated by the rest of rice leaves and rice-leaf disease boundaries should be not so differently changed and acceptable. Moreover, sometimes shadows unexpectedly occur and interfere brightness of rice leaves. These shadows affect in disease recognition and cause errors of area ratio. Shadow treatment should be considered in the preprocessing in the future.
Conclusion
Farmers who plant paddy rice in a big paddy rice field normally expect to install an automatic insecticide spray system for rice-leaf disease treatment in real time. To develop the automatic insecticide spray system, functions of surveillance, rice leaf detection, rice-leaf-disease detection and classification, severity level determination, etc. are required. Since multiple diseases sometimes appear on a rice leaf, the rice-leaf disease classification should be able to deal with multiple diseases in a leaf. To solve the research problem of multiple rice-leaf diseases in a rice leaf, this article therefore proposes a boundary-based rice-leaf-disease classification and severity level estimation methods for automatic insecticide injection. The method first segments and recognizes rice leaf in an image based on color feature, segment boundaries in a rice leaf as rice-leaf disease candidates and classify rice leaf diseases in those segmented candidates. In boundary-based classification, possibility of coarse-to-fine concept is proposed to enhance the classification accuracy. If all rice-leaf disease classes can be classified in the coarse step, the tolerance in the fine step is enlarged and accuracy is increased. Furthermore, severity level is proposed to determine area ratio between rice leaf and diseases. The performance of proposed rice-leaf-disease classification method and severity level estimation have been evaluated by experiments with 5,000 and 1,600 images of multiple rice leaves and diseases, respectively, and the results revealed 96.82% and 95.17%, accuracies on classification and severity level estimation, respectively.
Acknowledgments
The authors are thankful to expert Ms. Wanporn Khemmuk from Rice Department, Ministry of Agriculture and Cooperatives Thailand and thankful to Ph.D. scholarship supported by Rajamangala University of Technology Tawan-ok, (RMUTTO).
References
A.Hussein, M. (2017). Comparison of features extraction algorithms used in the. J. Pure Appl. Sci., 512-527.
Al-Amin, M., Karim, D. Z., & Bushra, T. A. (2019). Prediction of rice disease from leaves using deep convolution neural network towards a digital agricultural system. Proc. 22nd Int. Conf. Comput. Inf. Technol. (pp. 1-5). IEEE Xplore. doi:https://doi.org/10.1109/ICCIT48885.2019.9038229
Atrio, À. R., & Popescu-Belis, A. (2022). On the interaction of regularization factors in low-resource neural machine translation. Proc. 23rd Annu. Conf. Eur. Assoc. Mach. Transl., (pp. 111–120). Retrieved from https://aclanthology.org/2022.eamt-1.14
Bakar, M. N., Abdullah, A. H., Rahim, N. A., Yazid, H., Misman, S. N., & Masnan, M. J. (2018). Rice leaf blast disease detection using multi-level colour image thresholding. J. Telecommun. Electron. Comput. Eng., 10, 1-6. Retrieved from https://jtec.utem.edu.my/jtec/article/view/4036
Bari, B. S., Islam, M. N., Rashid, M., Hasan, M. J., Razman, M. A., Musa, R. M., . . . Majeed, A. P. (2021). A real-time approach of diagnosing rice leaf disease using deep learning-based faster R-CNN framework. PeerJ Comput. Sci., 7. doi:https://doi.org/10.7717/peerj-cs.432
Chettanachit, D., Nilpanit, N., Arunyanart, P., Sirisantana, W., Rattanakarn, W., Dhitikiattipong, R., . . . Arayapan, T. (2007). Rice disease and prevention. Bangkok: Rice Department.
Cunnac, S. (2021). Xanthomonas Oryzae Co-infections: Developing methodologies to address their epidemiological consequences. Agropolis Fondation.
Datacountry. (2021). Google Drive. Retrieved from https://drive.google.com/file/d/1IE53drYkIwe04N13RRr7bNZdb8100sLy/view
Dataset. (2021). Google Drive. Retrieved from https://drive.google.com/drive/folders/17dsYpjuwU23bGI7kbb_1A5Sd9BQHMsVU
eeNewsEurope. (2022). Resnet50 - A misleading machine learning inference benchmark for megapixel images. Retrieved from https://www.eenewseurope.com/en/resnet-50-a-misleading-machine-learning-inference-benchmark-for-megapixel-images/#:~:text=ResNet%2D50%20is%20a%20classification,typically%20measured%20with%20INT8%20operation
Feng, J., Yang, L., Yu, C., Di, C., & Li, G. (2020). Image recognition of four rice leaf diseases based on deep learning and support vector machine. Comput. Electron. Agric., 179. doi:https://doi.org/10.1016/j.compag.2020.105824
Ganesh, L. J., Raate, M. N., Nithin, T. N., Pavan, G., & Nithyashree, S. (2020). Precision agriculture robot for seeding function and leaf disease detection. Int. J. Eng. Res. Technol., 9(8), 40-43.
Ghosal, S., & Sarkar, K. (2020). Rice leaf diseases classification using CNN with transfer learning. Proc. 2020 IEEE Calcutta Conf. IEEE Xplore. doi:https://doi.org/10.1109/CALCON49167.2020.9106423
Gurevin, D., Bragin, M., Ding, C., Zhou, S., Pepin, L., Li, B., & Miao, F. (2021). Enabling retrain-free deep neural network pruning using surrogata lagrangian relaxation. Proc. 30th Int. Joint Conf. Artif. Intell., (pp. 2497–2504). doi:https://doi.org/10.24963/ijcai.2021/344
Hasan, M. J., Mahbub, S., Alom, M. S., & Nasim, M. A. (2019). Rice disease identification and classification by integrating support vector machine with deep convolutional neural network. Proc. 1st Int. Conf. Adv. Sci. Eng. Robot. Technol. (ICASERT). IEEE Xplore. doi:https://doi.org/10.1109/ICASERT.2019.8934568
International Rice Research Institute. (2002). Standard Evaluation System for Rice (SES). International Rice Research Institute. Retrieved from http://www.knowledgebank.irri.org/images/docs/rice-standard-evaluation-system.pdf
Jabir, B., & Falih, N. (2021). Dropout, a basic and effective regularization method for a deep learning model: A case study. Indonesian J. Electr. Eng. Comput. Sci., 24(2), 1009. doi:http://doi.org/10.11591/ijeecs.v24.i2.pp1009-1016
Jung, J. K. (2018). Keras cats dogs tutorial. Retrieved from https://www.jkjung-avt.github.io/keras-tutorial/
Latif, G., Abdelhamid, S. E., Mallouhy, R. E., Alghazo, J., & Kazimi, Z. A. (2022). Deep learning utilization in agriculture: Detection of rice plant diseases using an improved CNN model. Plants, 11(17), 2230. doi:https://doi.org/10.3390/plants11172230
Liang, W.-j., Zhang, H., Zhang, G.-f., & Cao, H.-x. (2019). Rice blast disease recognition using a deep convolutional neural network. Nature, 9. doi:https://doi.org/10.1038/s41598-019-38966-0
Loris, N., Gianluca, M., & Alessandra, L. (2021). Exploiting Adam-like optimization algorithms to improve the performance of convolutional neural networks. arXiv:2103.14689, 1-5. doi:https://doi.org/10.48550/arXiv.2103.14689
Matin, M. M., Khatun, A., Moazzam, M. G., & Uddin, M. S. (2020). An efficient disease detection technique of rice leaf using AlexNet. J. Comput. Commun., 8(12). doi:https://doi.org/10.4236/jcc.2020.812005
Mehta, A., & Masdekar, M. (2018). Precision agriculture - A modern approach to smart farming. Int. J. Sci. Eng. Res., 9(2), 23-26.
Mekha, P., & Teeyasuksaet, N. (2021). Image classification of rice leaf diseases using random forest algorithm. Proc. Joint Int. Conf. Digital Arts, Media Technol. ECTI Northern Section Conf. Electr. Electron. Comput. Telecommun. Eng. (pp. 165-169). IEEE Xplore. doi:https://doi.org/10.1109/ECTIDAMTNCON51128.2021.9425696
Muthayya, S., Sugimoto, J. D., Montgomery, S., & Maberly, G. F. (2014). An overview of global rice production, supply, trade, and consumption. Ann. N.Y. Acad. Sci., 1324(1), 7-14. doi:https://doi.org/10.1111/nyas.12540
Plantwise Knowledge Bank. (2013). PlantwisePlus Knowledge Bank. Centre for Agriculture and Biosciences International PlantwisePlus. Retrieved from https://plantwiseplusknowledgebank.org
Pothen, M. E., & Pai, M. L. (2020). Detection of rice leaf diseases using image processing. Proc. Fourth Int. Conf. Comput. Methodol. Commun. (pp. 424-430). IEEE Xplore. doi:https://doi.org/10.1109/ICCMC48092.2020.ICCMC-00080.
Prajapati, H. B., Shah, J. P., & Dabhi, V. K. (2017). Detection and classification of rice plant diseases. Intell. Decis. Technol., 11(3), 357-373. doi:https://doi.org/10.3233/IDT-170301
Ramesh, S., & Vydeki, D. (2020). Recognition and classification of paddy leaf diseases using Optimized Deep Neural network with Jaya algorithm. Inf. Process. Agric., 7(2), 249-260. doi:https://doi.org/10.1016/j.inpa.2019.09.002
Rice Knowledge Bank. (2013). Rice Knowledge Bank. International Rice Research Institute. Retrieved from http://www.knowledgebank.irri.org.
Ricepedia.org. (2013). Ricepedia. Retrieved from http://www.ricepedia.org/rice-as-food
Sethy, P. K., Barpanda, N. K., Rath, A. K., & Behera, S. K. (2020). Deep feature based rice leaf disease identification using support vector machine. Comput. Electron. Agric., 175. doi:https://doi.org/10.1016/j.compag.2020.105527
Shafi, U., Mumtaz, R., García-Nieto, J., Hassan, S. A., Zaidi, S. A., & Iqbal, N. (2019). Precision agriculture techniques and practices: From considerations to applications. Sensors, 19(17). doi:https://doi.org/10.3390/s19173796
Shrivastava, V. K., Pradhan, M. K., Minz, S., & Thaku, M. P. (2019). Rice plant disease classification using transfer learning of deep convolution neural network. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-3/W6. doi:https://doi.org/10.5194/isprs-archives-XLII-3-W6-631-2019
Tollenaere, C., Lacombe, S., Wonni, I., Barro, M., Ndougonna, C., Gnacko, F., . . . Brugidou, C. (2017). Virus-bacteria rice co-infection in Africa: Field estimation, reciprocal effects, molecular mechanisms, and evolutionary implications. Front. Plant Sci., 8. doi:https://doi.org/10.3389/fpls.2017.00645
Tosawadi, T., Kasetkasem, T., Laungnarutai, W., Phatrapomnant, T., & Kumazawa, I. (2021). Automatic rice plant disease evaluation method based on anomaly detection and deep learning. Proc. 18th Int. Conf. Electr. Eng./Electron. Comput. Telecommun. Inf. Technol. IEEE Xplore. doi:https://doi.org/10.1109/ECTI-CON51831.2021.9454737
Ullah, A., Ahmad, J., Muhammad, K., & Lee, M. Y. (2017). A survey on precision agriculture: Technologies and challenges. Proc. 3rd Int. Conf. Next Gener. Comput., (pp. 1-3).
Wikipedia. (2022). Six Sigma. Retrieved from https://en.wikipedia.org/wiki/Six_Sigma