Encyclopedia
Please note this is a comparison between Version 1 by Fraol Gelana Waldamichael and Version 4 by Peter Tang.
Cereals are an important and major source of the human diet. They constitute more than two-thirds of the world’s food source and cover more than 56% of the world’s cultivatable land. These important sources of food are affected by a variety of damaging diseases, causing significant loss in annual production. In this regard, detection of diseases at an early stage and quantification of the severity has acquired the urgent attention of researchers worldwide. One emerging and popular approach for this task is the utilization of machine learning techniques.
- cereal crop
- plant disease
- machine learning
- deep learning
1. Introduction
Advancements in the area of machine learning and computer vision in the past decade had had a profound effect on the utilization of machine learning techniques in different sectors [1]. Machine learning approaches are being used from the medical [2][3][4][5][6][7][8][9][2,3,4,5,6,7,8,9] to the security sector [10]. Recently, many works [11] have been undertaken on the application of machine learning in the agriculture sector for the detection of plant diseases, such as coffee [12] and Enset [13], Crop yield prediction [14], quality and growth monitoring [15][16][15,16], supply chain performance [17], and water stress determination [18].
Plants constitute 98% of the world’s diet, two-thirds of which are Cereals [19]. The eight major kinds of cereal, wheat, maize, rice, barley, sorghum, oats, millets, and rye cover 56 percent of the world’s arable land. Wheat, maize, and rice account for 80% of global cereal production [19]. Plant diseases are the major cause of global crop yield reduction, resulting in 10% loss of all the global food production [20]. The major plant disease-causing pathogens are viruses, bacteria, Oomycetes, fungi, nematodes, and other parasitic plants [20]. When infections occur to a large extent, losses to cereal crop production could reach as high as 50% [21]. Many laboratory techniques are available for the identification and detection of plant pathogens [20], but rapid and early detection is an important factor in the successful containment and control [22].
2. Cereal Crops and Diseases
Cereal is a crop closely related to grass and that is cultivated for its seed and is consumed as food by humans [23][29]. According to the Cereal Disease, Methodology Manual [19], the eight major kinds of cereal, covering 56% percent of the world’s arable land are Wheat, Maize, Rice, Barley, Sorghum, Oats, Millet, and Rye.2.1. Wheat
Wheat is the most dominant and important source of food for humans and livestock [24][30]. It is the main ingredient in flour, which is used in the making of bread, biscuits, and pastry [19]. Wheat is cultivated across all parts of the earth, from Russia in the northern hemisphere to Argentina in the south [24][30]. Diseases pose a serious threat to the global production of wheat [20]. Diseases on wheat are caused by a variety of pathogens. These are, Fungai, Viruses, Bacteria, Insects and Nematodes [25][31]. Some of the commonly occurring wheat diseases are given in Table 1.
, where barley adaptation to high altitude environments makes it an important source of food and beverages for millions of people [33][39]. Barley is affected by over 80 different diseases caused by a variety of pathogens [34][40]. Some of these are summarized in Table 4.
|
Pathogen |
Disease |
|---|---|
|
Leaf Rust (Brown Rust), Stem Rust (Black Rust), |
|
|
Stripe Rust (Yellow Rust), Common Root Rot, |
|
|
Fungus |
Common and Dwarf Bunt (Stinking Smut), |
|
Wheat Blast, |
|
|
Tan Spot |
|
|
Bacterial Stripe (Black Chaff), |
|
|
Bacteria |
Basal Glume Rot and Bacterial Leaf Blight, |
|
Pathogen |
Disease |
||
|---|---|---|---|
Disease |
|||
|
Fungus |
|||
|
Fungus |
Stripe rust, Leaf rust, Stem rust |
||
|
Anthracnose, Leaf blight, Zonate leaf spot |
Snow mold, Brown rust, Ergot |
Powdery mildew, Downy mildew |
|
|
Tar spot, Charcoal rot | |||
Bacterial Spike Blight (Gummosis) | |||
|
Eye spot, Sharp eyespot |
Rust, Gray leaf spot |
||
|
Bacteria |
Bacterial stripe |
||
|
Virus |
Streak disease |
Barley Yellow Dwarf, |
Table 6.
|
Pathogen |
Disease |
|---|---|
|
Fungus | |
|
Powdery mildew, Stem rust, Glume blotch |
|
|
Virus |
Yellow dwarf |
Virus | |
3. Machine Learning-Based Cereal Crop Disease Detection
3.1. Machine Learning in Wheat Disease Detection
Bao et al. [38][48] applied elliptical-maximum margin criterion metric learning to the identification and severity estimation of powdery mildew and stripe wheat disease types. The researchers choose the E-MMC algorithm since it is better suited to finding nonlinear transformations in patterns, and their results show that it achieved superior results when compared to the SVM algorithm. For testing their algorithm, the researchers prepared a dataset from farms around the province of Beijing. In total, they collected 360 images. Disease spot segmentation was performed by using the Otsu thresholding algorithm and feature extraction using HSV histogram, Color moments for color attributes, and LBP and Gabor for texture attributes.
Identification of various wheat diseases using hyper-spectral image data were performed by [39][40][41][42][47,58,59,60]. Identification of wheat powdery mildew disease using linear regression and an SVM (Figure 1) classifier on hyper-spectral data ranging from 656 nm to 784 nm was implemented by Huang et al. [40][58]. The authors employed the Relief-F algorithm to identify the best spectral bands and evaluation of the SVM algorithm was performed by k-fold cross-validation. In addition, Huang et al. [41][59] proposed an SVM-based detection of Fusarium Head Blight on wheat heads using hyperspectral imagery. Here, Fishers Linear Discrimination (FLD) was implemented for dimensionality reduction. An in-field detection of yellow rust and fusarium head blight in wheat-based on the ground and UAV-based platforms was discussed by Bohnenkamp et al. [43][61] (Figure 2) and Xiao et al. [44][62].
Figure 1. Flow chart for hyper-spectral image data analysis and processing for wheat rust detection [40].
Summary of various wheat leaf disease datasets is presented in Table 8.
Table 8.
Performance comparison of selected studies on machine learning based wheat disease detection and corresponding datasets.
|
Citatation |
Year |
Data Type |
# of Classes |
Sample Size |
# of Classes Method |
Sample Size Accuracy % |
|
|---|---|---|---|---|---|---|---|
Method | Accuracy % | ||||||
|
Crown rust, Stem rust, Powdery mildew |
|||||||
2021 |
Image |
12 |
1100 |
MobileNetV2 |
99.67% |
||
|
Wang et al. [ | |||||||
2020 | |||||||
Image | 11 |
5932 |
SVM |
98.38% |
|||
|
2019 |
Image |
3 |
7448 |
faster R-CNN |
96.21% |
3.3. Machine Learning in Maize Disease Detection
An Enhanced CNN for the detection of nine classes of maize leaf disease was proposed by Agarwal et al. [63][75]. They proposed a convolutional neural network with receptive field enlargement to enhance the feature extraction performance of the CNN, which is required due to the complexity of maize leaf images. To accomplish this task, the authors collected a dataset of 500 images of maize leaves belonging to nine different classes of maize leaf disease at different stages. The performance of the proposed approach was compared to existing models like AlexNet and GoogleNet and provided an improved classification accuracy of 95.12%. Sibiya et al. [64][76] developed a convolutional neural network for the detection of three different maize leaf diseases by using the Neuroph framework for the java programming language. The proposed approach gave a classification accuracy of 93.5%.
Barman et al. [65][77] proposed a MobileNet architecture-based maize leaf disease detection that will be deployed on Android mobile devices. The authors utilized a transfer learning approach to fine-tune the pre-trained MobileNet architecture. For this task, they used a public dataset (PlantVillage) with a total of 3852 images of four different classes of maize leaf diseases. The proposed approach yielded an accuracy of 94.53%.
Hasan et al. [66][78] proposed a hybrid network by combining a convolutional neural network and bi-directional LSTM for the detection of nine classes of maize leaf diseases. bi-LSTM was selected by the authors to better accelerate CNN’s classification accuracy and increase the co-relation among extracted features. Training of the model was performed on the PlantVillage dataset, which contains 2500 images of maize leaves affected by nine different types of diseases. They implemented various image augmentation techniques and increased the size of the dataset to 29,065 images. The proposed approach achieved a classification accuracy of 99.02%, exceeding existing deep learning methods.
Xu et al. [67][79] proposed a multi-scale convolutional global pooling convolutional neural network based on the AlexNet and Inception architecture. The proposed model improves on the AlexNet architecture by replacing the last fully connected layer with a global pooling layer and adding a batch normalization layer. This is implemented to solve the low accuracy achieved and the large training data size required when utilizing transfer learning. Training and testing of the proposed model were performed on the PlantVillage dataset. The authors found that the proposed approach improves average precision by more than 2% when compared to AlexNet. A VGG16 deep learning architecture-based maize disease identification was proposed by Tian [68][80]. In this work, a transfer learning approach was used to fine-tune the pre-trained VGG16 architecture on a dataset consisting of 7858 images of maize leaves affected by six types of diseases. The proposed method achieved a classification accuracy of 96.8%. Summary of various maize leaf disease datasets is presented in Table 10.
Table 10.
Performance comparison of selected studies on machine learning based maize disease detection and corresponding datasets.
|
Citatation |
Year |
Data Type |
# of Classes |
Sample Size |
Method |
Accuracy % |
|||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
2021 |
Image |
9 |
500 |
CNN |
95.12% |
||||||||||
Smut disease, Leaf blight |
|||||||||||||||
|
2021 |
Image |
3 |
2370 |
] |
MobileNetV2 |
2019 |
Image (PlantVillage) |
9 |
94.65% |
||||||
2500 | CNN |
95.5% |
Net blotch, Spot blotch, Stripe disease |
||||||||||||
2019 |
2021 Image |
Image (PlantVillage) 1 |
9 5808 |
2500 CNN |
MobileNetV2 95.83% |
||||||||||
93.5% |
Bacteria |
||||||||||||||
|
Bacterial blight |
|||||||||||||||
2021 |
Image |
] |
3 |
2020 |
300 |
CNN |
90% |
Image (PlantVillage) |
9 |
2500 |
LSTM |
99.02% |
Virus |
||
|
2021 |
] |
Image |
3 |
1500 |
2021 VGG19 |
||||||||||
|
Xu et al. [67][ |
Barley stripe mosaic |
||||||||||||||
| Cereal tillering virus | Barley Stripe Mosaic, |
||||||||||||||
|
Wheat Streak Mosaic |
|||||||||||||||
|
Aphids, Stink Bugs, |
|||||||||||||||
|
Cereail Leaf Beetle, |
|||||||||||||||
|
Insect |
Thrips, |
||||||||||||||
|
Hessian Fly, Wireworms, |
|||||||||||||||
|
Mites |
|||||||||||||||
|
Nematode |
Seed Gall Nematode |
||||||||||||||
|
Cereal Cyst Nematode |
|||||||||||||||
|
Root Knot Nematode |
|||||||||||||||
|
Root Lesion Nematode |
2.2. Maize (Corn)
Maize is an important staple food crop that is grown all over the globe. It is the largest grown cereal per unit area, yielding 785 million tons annually [26][32]. Besides being a source of food, maize, and its products are used as raw materials for many industrial applications. Maize is prone to many types of diseases caused by a variety of pathogens. Fungal pathogens are the major causes of maize disease, while bacterial and viral diseases are less common but pose a serious threat [21][27][21,33]. Commonly occurring maize diseases are given in Table 2.
|
Pathogen |
Disease |
|---|---|
|
Fungus |
Gray leaf spot, Brown spot, |
|
Stripe Rust (Yellow Rust) |
|
|
Common rust, Smut, |
|
|
Northernl eaf blight, Southern leaf blight |
|
|
Bacteria |
Corn stunt disease |
|
Stewart wilt |
|
|
Bacterial stalk rot |
|
|
Bacterial leaf strip |
|
|
Virus |
Leaf fleek |
|
Mosaic |
|
|
Yellow dwarf |
2.3. Rice
Rice is the second most-produced cereal crop in the world [28][34]. It is the main source of food for billions of people in the world and is one of the primary food sources for the majority of people in Asia [29][35] with around 500 metric tons [30][36] of rice milled every year. Rice is susceptible to a variety of disease-causing pathogens that attack the leaf, the seed, the stem, and the root [31][37], some are given in Table 3.
2.5. Sorghum
Sorghum is the fifth most important cereal crop after wheat, maize, rice, and barley [35][41]. It is cultivated around the globe and used as a 0 source of food and energy, when used as a bio-fuel [36][42]. Sorghum production is highly affected by fungal and viral diseases, at times causing around 28% loss in production [37][43]. Some commonly occurring sorghum diseases are presented in Table 5, Table 6 and Table 7.
2.4. Barley
Barley is an important staple food cereal crop, although it is produced in much less quantity than wheat, maize, and rice [19]. It is farmed in significant quantities in sub-Saharan countries like Ethiopia [32][38]
Table 7.
|
Pathogen |
Disease |
||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
Fungus |
|||||||||||||||||
2021 | |||||||||||||||||
Image |
1 |
450 |
CNN |
89.9% |
|||||||||||||
|
Tagel et al. | Image |
2021 |
Image (PlantVillage) 3 |
9 |
2500 |
99.38% |
|||||||||||
276 | CNN |
TCI-ALEXN 91.47% |
99.18% |
||||||||||||||
|
2018 |
Image |
||||||||||||||||
|
Tian [68 |
2020 | 4 |
8828 |
AlexNet |
84.54% |
||||||||||||
][ | Image |
15 |
80] |
2019 | 500 |
Image (PlantVillage) | DenseNet |
9 |
2500 | 94.07% |
VGG16 |
96.8% |
|||||
|
2017 |
Image |
6 |
9230 |
2019 |
VGG-FCN |
Image |
97.95% |
||||||||||
2 | 284 |
InceptionV3 |
99% |
2019 |
Hyper-spectral |
2 |
|||||||||||
284 | v-SVR |
2019 | 0.99R |
Image | 2 |
||||||||||||
9 | 1080 |
InceptionV3 |
97.5% |
2019 |
Hyper-spectral |
2 |
145 |
Linear Regression |
0.75R2 |
||||||||
|
2019 |
Hyper-spectral |
2 |
89 |
SVM |
85.7% |
3.2. Machine Learning in Rice Disease Detection
Identification and classification of 12 types of rice leaf diseases using MobileNetV2 architecture and attention mechanism were proposed by Chen et al. [52][63]. The MobileNetV2 architecture was pre-trained on the ImageNet dataset and fine-tuned by using the transfer learning approach on a smaller local dataset. The authors utilized Channel Attention Mechanism (CAM) to better learn the inter-channel relationships. For fine-tuning and testing their proposed model, the authors collected a total of 1100 images of healthy and disease rice leaves. These 660 were compiled from various sources on the internet and 440 were collected from the field. The proposed model achieved an average classification accuracy of 99.67%. Similarly, Wang et al. [53][64] proposed a MobileNetv2 based approach for the classification of three types of rice leaf diseases by utilizing attention mechanism and Bayesian optimization. Model training and validation were performed on a public dataset of 2370 images belonging to three classes of rice disease and one healthy class. The authors achieved a classification accuracy of 94.65%.
Liang et al. [54][65] proposed a convolutional neural network-based rice blast disease detection approach. The authors proposed two CNN architectures, the first network containing four convolutional layers, four max-pooling layers, and three fully connected layers, and ReLU after each layer (Figure 3a) and a second network having the same convolutional layers and max-pooling layer structure as the first network, but with two additional fully connected layers as shown in (Figure 3b). The two models were trained on a custom dataset of 5808 images of healthy and rice blast infected leaves. The dataset was collected on-site and is divided into 2906 positive (rice blast infected) and 2902 healthy images. The authors utilized 5-fold cross-validation and a selected the second model due to its inherent stability on small datasets and chieved an accuracy of 95.83%. The proposed approach was compared to hand-crafted approaches like Local Binary Patterns Histogram (LBPH), Haar-WT. The comparison result suggests that the proposed CNN method achieves superior feature extraction and classification results. A similar approach for the detection and classification of three classes of rice disease was proposed by Rahman et al. [55][66]. The authors proposed a convolutional neural network trained on a dataset of 300 images containing three types of rice leaf disease (Brown spot, Leaf blight, and Hispa) and one healthy class. The model achieved a classification accuracy of 90%. This low classification accuracy is a result of the small dataset size the authors used and the lack of utilizing transfer learning. Ramesh et al. [56][67] proposed a convolutional neural network approach for the detection of three classes of rice disease. The authors utilized HSV color space for the separation of background and foreground and the K-means algorithm for disease segmentation.
A random forest classifier for the detection and classification of three types of rice leaf disease was proposed by Saha and Ahsan [57][68]. A local dataset compromising a total of 276 images of healthy and infected rice leaves was collected by the authors for testing and training their proposed algorithm. Feature extraction was implemented by using intensity moments. The proposed approach achieved a classification accuracy of 91.47%. A deep learning method for the detection of 15 different rice diseases was implemented by Chen et al. [58][69]. The authors developed a deep learning architecture based on the fusion of existing DenseNet and Inception architectures. For testing the proposed model, the authors compiled a dataset consisting of 500 images belonging to 15 classes of rice disease. Their proposed model achieved a classification accuracy of 94.07%.
