Image recognition of four rice leaf diseases based on deep learning and support vector machine

Highlights

• This study uses deep learning and SVM approach to identify four rice diseases.

• The CNNs includes 7 layers and the optimal parameters of SVM are obtained by experiment.

• The shape and color feature of lesion images are extracted by manual methods and CNNs.

• The higher recognition performance of CNN plus SVM with optimal parameters is achieved.

Abstract

In the field of agricultural information, identification and prediction of rice leaf diseases has always been a research focus. Deep learning and support vector machine (SVM) technology are hot research topics in the field of pattern recognition at present. Their combination can not only solve the problem effectively, but also improve the recognition accuracy. In this study, firstly, we use convolution neural networks (CNNs) to extract the rice leaf disease images features. Then the SVM method is applied to classify and predict the specific disease. The optimal parameters of SVM model are obtained through the 10-fold cross validation method. The experimental results show that when the penalty parameter $C=1$ and the kernel parameter g = 50, the average correct recognition rate of the rice disease recognition model based on deep learning and SVM is 96.8%. This accuracy is higher than that of the traditional back propagation neural networks models. This study provides a new method for the further research of crop diseases diagnosis by using deep learning.

Introduction

Rice, wheat and maize are the three major food crops in the world (Li, 2004). Among them, rice has the most sown area, the largest total yield and the highest yield per unit in China which has become one of the main grains of Chinese people. China’s cultivated area accounts for about a quarter of the country’s arable land, and the annual output accounts for half of the country’s total grain output. For more details, we refer readers to the paper (Kavya et al., 2016, Pragati and Surekha, 2017, Yi and Deng, 2017, Shen et al., 2019, Ma and Wang, 2004, Lu et al., 2017). It has been estimated that there will be more than nine billion people in the world by 2050 (Godfray et al., 2010). Nevertheless, rice diseases have been seriously affecting rice production. Consequently, it is a big challenge for the agricultural community to ensure the food security of such a large population. During the growth of rice, many diseases are often accompanied, such as rice blast, sheath blight, jute leaf spot, rice curl disease, bakanae disease, bacterial leaf blight, and bacterial leaf streak disease. These diseases occur in every part of rice, such as leaf, neck, and ear. Among them, rice blast, red blight, stripe blight and sheath blight are the four diseases with the largest occurrence probability and the widest influence.

Diseases will not only lead to production reduction, but also bring environmental pollution. When rice is attacked by diseases, they will seriously affect the yield of crops. Among them, a loss of 10–15% in rice production are caused by rice plant diseases (Peng et al., 2009). In serious cases, it can reach 40–50% or even no income, which will bring huge economic losses to the farmers who have worked hard for a year. However, farmers generally do not know enough about the diseases of rice or even crops, so they can not fully understand the occurrence of each disease and the specific characteristics of the disease spots, and thus can not judge the disease and apply medicine in time. As a result, too much, too little or even wrong medicine will be used, which will delay the treatment of disease and damage the soil and the environment. Therefore, it is very important to find the disease and identify the disease type as soon as possible.

In general, detection of rice plant diseases is based either on the visual assessment of the symptoms or the experimental result by culturing pathogens in the laboratory. The visual assessment is a subjective approach and prone to error. Whereas, culturing the pathogens in the laboratory is a time-consuming process and may not provide the result in time(Barbedo, 2013). Besides these limitations, both the conventional approaches require experts to identify the diseases and it is difficult for farmers to get access to experts due to the interior region of their agriculture field. These issues have prompted the research community to investigate various algorithms and develop automatic methods to detect and classify rice plant diseases, and simultaneously motivates the farmers to decide and choose the right pesticides.

With the development of computer and internet technology, recognition methods based on image processing and pattern recognition technology constantly emerge (Lai et al., 2009, Rahul et al., 2016, Kavya et al., 2016, Li et al., 2013, Pragati and Surekha, 2017). The main research fields include rice disease image acquisition methods, disease image preprocessing methods, feature extraction and optimization of disease spot image, and establishment of rice disease automatic recognition modeled by Bayesian classifier, back propagation neural network, support vector machine (SVM) and other methods (Tharwat et al., 2018, Dey et al., 2017, Guan, 2018, Qiu et al., 2019).

In recent years, digital image processing technology is becoming more and more mature and has been widely used in many fields which promotes the new development of these disciplines. The applications in the field of agriculture mainly includes the identification and diagnosis of crop diseases, pests and weeds (Burks et al., 2000, Geea et al., 2008, Sgaard, 2005), the identification and diagnosis of crop deficiency disease (Mao et al., 2005), the classification and inspection of crop seed quality (Zayas and Flinn, 1998, Steenhoek et al., 2001), the quality detection and classification of agricultural products (Fang et al., 2004), and some results have been achieved. Image processing technology is widely used in agriculture, especially in crop disease recognition which mainly includes the identification of leaf diseases of corn, rice, eggplant and other crops, fruit diseases of cucumber, tomato and other fruits, and common diseases of cashmere cotton, sunflower, tea, tobacco and other economic crops.

At present, many experts and scholars have used the technology of deep learning and SVM to identify and study the rice and general plant diseases. Ze-xin Guan (Guan et al., 2010) has used the image processing technology and the pattern recognition technology to respectively segment the rice disease images and extract the classify features, and a rice disease recognition system has been established. Through the analysis of the experimental results, it is concluded that this system has a better recognition performance. You-wen Tian et al. (Tian and Li, 2006) have studied the recognition of cucumber and corn diseases by using SVM, and have achieved good recognition results. Taking the wheat total erosion disease as the research object, Fu Wei (Fu, 2015) has used the unmanned aerial vehicle to take the wheat disease images and carried on the pretreatment and the characteristic parameter extraction of the collected disease images. Then, the disease degree of wheat total erosion disease has been obtained by comparing the color characteristic parameters. R. Pydipati et al. (Pydipati et al., 2006) have employed color co-occurrence method to determine texture-based hues. Saturation and intensity (HSI) color features combined with statistical classification algorithm have been used to identify diseases and normal citrus leaves under laboratory conditions. Yun-lan Tan et al. (Tan et al., 2019) have used the deep convolution network to establish the rice disease identification model to classify and identify eight kinds of rice diseases, of which the accuracy of sheath blight is 93%. Bo et al. have taken the image of vegetable and fruit diseases as the research object, then the image processing technology, the deep learning method and the convolution neural network (CNN) have been respectively used to preprocess the image, improve the recognition rate of the vegetable and fruit image and build the recognition model so as to accomplish the image recognition task (Bo et al., 2018). Shen Wei Zheng et al. (Shen et al., 2019) have studied the automatic identification method of rice blast, sheath blight and bacterial leaf blight based on BP neural network. The parameters have been optimized by using the single factor analysis of variance and the effect of BP neural network model. The results have shown that the model has a good recognition performance.

Although the main methods used in the above studies include some traditional analysis methods, most of them still use a combination of multiple methods or propose a new research method. In pattern recognition, linear classifier based on statistics (Yusako et al., 2001a), Bayesian decision theory, artificial neural network (Pydipati et al., 2005, El-Faki et al., 2000, Ma and Wang, 2004), fuzzy recognition technology (Mohamed et al., 2006, Vioix et al., 2004, Qiu et al., 2002) and SVM (Tian et al., 2006) are widely used. At present, rice disease recognition based on digital image has become a research hotspot (Barbedo, 2013). Recently, CNN has also been applied to rice plant disease classification (Lu et al., 2017). Yong-gang Shi et al. (Shi et al., 2018) have proposed a segmentation method based on CNN and SVM to get the segmentation of hippo image. They use SVM to replace the output layer of CNN. The model can automatically extract image block features by training deep-seated network, then use the extracted image features to train SVM to improve image classification. The experimental results show that the segmentation accuracy of each region is greatly improved. In this study, in order to extract the effective feature parameters of rice disease images and establish the foundation of disease recognition system, we use mean shift method to separate the normal and abnormal parts of leaves so as to simplify the difficulty of image processing and further improve the processing speed. The motivation of using the mean-shift method as image segmentation method is that the area of the extracted spots is larger than that of the artificial threshold extraction. However, the existing image segmentation technology based on mean shift algorithm is not perfect, and the segmentation results are not stable. Therefore, a disease recognition model combining CNN and SVM is proposed to improve the recognition accuracy.

There are two main motivations to study the image recognition models of four rice diseases. (1) So far, this is still a challenge, and is becoming the main focus of rice disease recognition research. In order to obtain satisfactory disease recognition rate, this paper applies CNN and SVM to rice disease recognition. (2) According to the difference of rice disease spots, the mean shift algorithm and CNN are used to extract the optimal feature parameters.

There are two main contributions of this paper. (1) Aiming at the identification of four rice diseases, a new algorithm which combines CNN with SVM proposed in this paper. It is worth noting that the model can correctly and effectively identify rice diseases. (2) The experimental results show that the best combination of features and the best parameters of the model can not only improve the convergence speed, but also have higher recognition accuracy than the classical models such as feature extraction and disease recognition directly using CNN.