1. Introduction
Rice is one of the most important food crops for mankind, with a long history of cultivation and consumption. Half of the world’s population eats rice, mainly in Asia, southern Europe, tropical America and parts of Africa [
1,
2]. Rice sheath blight, caused by
Rhizoctonia solani, is one of the main diseases in rice production [
3]. Over the past 20 years, the occurrence of sheath blight in rice areas in China has gradually increased. The disease mainly occurs in leaf sheaths and leaves. It can occur over the whole growth period of rice and has a great impact on yield [
4,
5]. If the problem persists, it will have a huge impact on production, with estimates of a general reduction of 10% to 30%, and even a serious reduction of 50%. In the early stage of the disease, an oval, dark-green, water-stained lesion appears on the leaf sheath, and then gradually expands into a moiré shape with a grayish white in the middle [
6,
7]. Traditionally, agricultural practitioners recognize rice sheath blight disease using their naked eye to judge the health of the plant. This recognition method not only requires a lot of manpower, but also has low accuracy. During the recognition process, agricultural practitioners are required to go to the paddy field for field observation. Traditional disease recognition methods are time consuming, labor intensive, have low efficiency, and involve subjective judgments, which cannot meet the standards of real-time and rapid monitoring required by modern society [
8]. Therefore, the application of computer technology and image processing technology in the recognition and detection of rice sheath blight is of great value to rice production because it is fast, accurate and works in real time.
With the development of computer image recognition and deep learning technology in recent years, there are many studies on the recognition and classification of crop diseases around the world. Rozaqi [
9] et al. found a way to recognize the early blight and late blight of potato leaves, using deep learning training in combination with a convolutional neural network architecture. They had a ratio of the training set against the validation set of 7:3, and used 20 batch-processed images over 10 time periods. The final accuracy rate was 92%. Addressing apple gray spot and cedar rust, Li [
10] et al. used image segmentation methods to compare and improve the three convolutional neural network models of the support vector machine (SVM), ResNet and VGG. Finally, they chose the ResNet-18 deep learning model and achieved a higher recognition rate. Almadhor [
11] et al. proposed an artificial intelligence-driven (AI-driven) framework to detect and classify the most common guava plant diseases. Full-name, 4E-color difference image segmentation was employed to segregate the infected areas. Color (RGB, HSV), histogram and texture (LBP) features were used to extract feature vectors, and advanced machine learning classifiers, such as fine KNN, complex tree and boosted tree, were used for recognition. The classification accuracy was 99%. Moreover, Oyewola [
12] et al. developed a novel deep residual convolution neural network (DRNN) for cassava mosaic disease (CMD) detection in cassava leaf images, which utilized a balanced image dataset to improve the classification accuracy. The DRNN model achieved an accuracy rate of up to 96.75%. In addition, Abayomi-Alli [
13] et al. studied the recognition of cassava leaf disease using a novel image color histogram transformation technology for data augmentation in image classification tasks, in order to solve the problem of using neural networks when one only has access to low-quality images. Four methods—resolution down-sampling, Gaussian blurring, motion blur, and overexposure—were carefully carried out for verification. Furthermore, Kundu [
14] et al. proposed a ‘Custom-Net’ model to identify blast and rice rust on pearl millet, and obtained images in real time through the ‘Automatic and Intelligent Data Collector and Classifier’ framework. The Custom-Net model had a classification accuracy of 98.78%. Finally, Hu [
15] et al. proposed a convolution neural network based on data augmentation and transfer learning to efficiently recognize corn leaf disease models. By simply tuning the GoogleNet pretraining network and controllably adjusting the parameters, including the optimizer and learning rate, the average recognition accuracy of corn disease, which consists of corn common leaf rust, corn common rust and corn northern leaf blight, was greater than 95%.
Although image recognition technology has been widely used in the field of agriculture, due to the planting characteristics of rice and the long incidence cycle of sheath blight, image recognition technology is relatively less used in rice disease recognition, especially for rice sheath blight. Until now, little direct agricultural work was reported in studies on rice disease recognition. Phadikar [
16] et al. used image segmentation technology for preprocessing operations, then employed a neural network as a classifier for classification operations and proposed a system for detecting rice disease images. Anthonys [
17] et al. selected rice blast and brown spot as diseases that could be targeted in the development of recognition and classification processes. In the recognition process, the digital morphology method was conducted to preprocess the rice disease images, and the membership function was utilized to classify 50 sample images. The recognition accuracy reached 70%, and a working system which can recognize the types of rice diseases was thus developed. In addition, Majid [
18] et al. used the fuzzy entropy method for feature extraction, employing a probabilistic neural network to recognize rice disease images, and developing a rice disease recognition application. Furthermore, Suman [
19] et al. performed histogram preprocessing on leaf images of rice blast and Cercospora leaf spot, and used SVM for classification and recognition. To conclude, rice stripe disease recognition methods based on image recognition technology and neural network technology has become the focus of significant present and future research.
In this study, a novel and facile method based on neural networks was rationally designed for the recognition and detection of rice sheath blight. Firstly, given the quality of real rice sheath blight images, preprocessing was carried out, including image smoothing, image enhancement and image segmentation, to greatly weaken the influence of complex backgrounds on the overall recognition. Secondly, the color and texture features of the rice sheath blight images were extracted to create the parameters that defined the subsequent training. Then, a backpropagation (BP) neural network model was built in MATLAB: 480 pictures were used as the training set, 120 pictures were used as the verification set and the input matrix and output matrix were established for training and testing. Finally, based on the actual recognition situation, we show how the number of hidden layer nodes of BP neural network can be efficiently adjusted for optimization. In China, few studies have been conducted so far that concern the recognition and detection of rice sheath blight using computer image recognition technology. The BP neural network model proposed in this paper provides a new platform for the recognition of rice sheath blight that is fast, accurate and works in real time. This will thereby promote the rapid development of sustainable, green and automatic agriculture.
The rest of the paper is arranged as follows:
Section 2 introduces the proposed method for processing the image of rice sheath blight.
Section 3 presents the experiment and results. The discussion on the proposed method is presented in
Section 4.
Section 5 concludes this paper.
3. Results and Discussion
Regarding the configuration of the hardware, we used an AMD Ryzen 7 4800H with 16 GB RAM and an NVIDIA GeForce RTX 2060, and the software is Matlab 2020a.
3.1. Data Collection
A total of 230 photos of rice with sheath blight were taken under natural light and saved in JPG format. In order to preserve the main diseased spots without compromising the integrity of the image, the image has been cropped to a size of 50 × 50 pixels. A total of 600 sample pictures have been obtained. Representative sample pictures are shown in
Figure 3.
3.2. Preprocessing and Feature Extraction
As shown in
Figure 4, after the experiment, the medfilt2 operator was found to have the best median filtering effect. Then the histeq operator was used to equalize the histogram of the image. Finally, by comparing the edge segmentation algorithm, Otsu method and region segmentation algorithm, it was found that the effect of edge segmentation using the Sobel operator is the best.
Figure 5 shows the feature extraction in the RGB color space for the facilitation of the recognition of rice sheath blight according to the color information. According to
Table 1, R is 97.8240, G is 88.8168 and B is 45.9972 in the first order moment of the diseased spots picture. In addition, the red component has the highest value and the blue component has the lowest, indicating that red is the most obvious color in the diseased spots picture.
Figure 4.
Image during preprocessing: (a) the original image; (b) median filter image; (c) histogram equalization image; (d) edge segmentation image.
Figure 4.
Image during preprocessing: (a) the original image; (b) median filter image; (c) histogram equalization image; (d) edge segmentation image.
Figure 5.
RGB component and histogram of rice sheath blight sample image: (a) red component; (b) green component; (c) blue component; (d) histogram of red, green and blue components.
Figure 5.
RGB component and histogram of rice sheath blight sample image: (a) red component; (b) green component; (c) blue component; (d) histogram of red, green and blue components.
Table 1.
Color moment of rice sheath blight.
Table 1.
Color moment of rice sheath blight.
Label | R | G | B |
---|
Mean value of first moment | 97.8240 | 88.8168 | 45.9972 |
Variance of second moment | 41.1104 | 34.4180 | 26.0005 |
The results of the gray difference statistical features comparison of leaves and diseased spots are shown in
Table 2.
It can be seen from
Table 2 that both the mean value (Mean) and entropy (Ent) of healthy leaves are smaller than those of diseased spots, which indicates that the texture of diseased spots is relatively rough. In addition, the differences between the mean values (Mean) and contrast ratios (Con) of the two is large.
3.3. Sample Training and Testing of the BP Neural Network
The number of the BP neural network’s input elements is equal to the dimension of the feature vector of the recognition object. The resolution of each picture is 50 × 50, with 2500 points. For each picture’s color characteristics (R, G and B), each picture has 7500 characteristics. For the three texture features that have been extracted from each image, namely, the mean value (Mean), contrast ratio (Con) and entropy (Ent), each image has three features. In summary, each picture has a total of 7503 features. From among the 600 pictures of rice, 480 pictures have been used as the training set to form a 7503 × 480 input matrix. The other 120 pictures are used as the verification set to form a 7503 × 120 verification input matrix. According to the input matrix, a 1 × 480 output matrix and a 1 × 120 verification output matrix are established. According to Equation (6), the number of hidden layer nodes is approximately 90.
Table 3,
Table 4 and
Table 5 show the experimental results when the number of hidden layer nodes is 80, 90 and 100, respectively.
When the number of hidden layer nodes for the neural network training was set at 80, around 100 samples out of the 120 test samples were successfully recognized in the five experiments that were conducted under the same conditions, with a recognition rate of 83.5%.
Table 4.
Test results with 90 hidden layer nodes.
Table 4.
Test results with 90 hidden layer nodes.
Label | Number of Samples | Recognition Quantity | Recognition Rates (%) |
---|
1 | 120 | 106 | 88.3 |
2 | 120 | 103 | 85.8 |
3 | 120 | 102 | 85.0 |
4 | 120 | 101 | 84.2 |
5 | 120 | 103 | 85.8 |
Mean value | \ | \ | 85.8 |
When the number of hidden layer nodes for the neural network training was set at 90, around 103 samples out of the 120 test samples were successfully recognized in the five experiments that were conducted under the same conditions, with recognition rate of 85.8%.
Table 5.
Test results with 100 hidden layer nodes.
Table 5.
Test results with 100 hidden layer nodes.
Label | Number of Samples | Recognition Quantity | Recognition Rates (%) |
---|
1 | 120 | 104 | 86.7 |
2 | 120 | 98 | 81.7 |
3 | 120 | 102 | 85.0 |
4 | 120 | 100 | 83.3 |
5 | 120 | 104 | 86.7 |
Mean value | \ | \ | 84.7 |
When the number of hidden layer nodes for the neural network training was set at 100, around 102 samples out of the 120 test samples were successfully recognized in the five experiments that were conducted under the same conditions, with a recognition rate of 84.7%.
Generally, the recognition rates of the three hidden layer nodes are almost the same. When the number of hidden layer nodes is 90, the recognition rate is the highest, which is 2.3% higher than that when the number of hidden layer nodes is 80 and 1.1% higher than that when the number of hidden layer nodes is 100. It can be concluded that with the increase in the number of hidden layer nodes, before the critical point, the memory and learning ability of the network can be enhanced, and the training recognition rate improves. After the critical point, as the number of hidden layer nodes increases, the BP neural network’s learning ability and recognition ability declines, and the induction ability and training recognition accuracy decrease. Therefore, the number of hidden layer nodes of BP neural network is finally set to 90, with 85.8% accuracy.
4. Discussion
Table 6 shows the confusion matrix for the detection results of 120 pictures (96 sheath blight pictures and 24 healthy rice pictures) in the validation set when the number of hidden layer nodes is 90. It can be seen that the image recognition accuracy of rice sheath blight is about 88%. The correct recognition rate of healthy rice pictures is relatively low, at about 75%. There may be two reasons for this. Firstly, there are only 24 pictures of healthy rice in the verification set. In the actual verification process, mistakes with one or two wrong pictures will therefore lead to relatively large errors. Secondly, some healthy rice also has traces or spots similar to grain blight, which the neural network is exposed to in the training process, resulting in inevitable errors occurring in the final recognition.
In
Figure 6, the second group of confusion matrices, whose recognition rate is close to the average value, is selected, the ROC plot is drawn and the AUC value is calculated. The ordinate is the true positive rate, which represents the ratio between the predicted number of positive samples and the actual number of positive samples, while the abscissa is the false positive rate, which represents the ratio between the predicted number of negative samples and the actual number of negative samples. The closer the curve is to the upper left corner, the larger the area formed by the curve and the horizontal axis, indicating that the recognition ability of the classification method is stronger. It can be seen that the method proposed in this paper has good classification ability for rice sheath blight images.
Table 6.
Confusion matrix for detection results.
Table 6.
Confusion matrix for detection results.
Label | Actual Class | Predicting Class |
---|
Sheath Blight Picture | Healthy Rice Picture |
---|
1 | Sheath blight picture | 87 | 9 |
Healthy rice picture | 5 | 19 |
2 | Sheath blight picture | 83 | 13 |
Healthy rice picture | 4 | 20 |
3 | Sheath blight picture | 85 | 11 |
Healthy rice picture | 7 | 17 |
4 | Sheath blight picture | 84 | 12 |
Healthy rice picture | 7 | 17 |
5 | Sheath blight picture | 85 | 11 |
Healthy rice picture | 6 | 18 |
This study has therefore realized the recognition and judgment of rice sheath blight images based on a BP neural network. The recognition accuracy can reach 85.8%, which is superior to that of the traditional manual recognition method. In other plant disease recognition experiments that used a neural network, the recognition rate of early blight and late blight of potato leaves reached 92% [
9], the detection rate of cassava mosaic disease was 96.75% [
12], and the average recognition accuracy of corn common leaf rust, corn common rust and corn northern leaf blight was more than 95% [
15]. In this research, there are several points that need to be paid attention to. The first point is that the number of pictures collected is not enough. The recognition rate can be enhanced by continuously increasing the number of pictures in the training set. Another point is that the disease spot of rice sheath blight is complex, changeable and without a fixed shape, which creates some difficulties for the recognition process. Further optimization of the BP neural network, such as improving the system transfer function and the network structure optimization, will significantly improve its ability to recognize crop diseases.