Intelligent Diagnose System of Wheat Diseases Based on Android Phone

Journal of Information & Computational Science 12:18 (2015) 6845 6852 December 10, 2015 Available at http://www.joics.com Intelligent Diagnose System of Wheat Diseases Based on Android Phone Yongquan Xia, Yaobin Li, Chen Li College of Computer and Communication Engineering, Zhengzhou University of Light Industry Zhengzhou 450001, China Abstract To improve wheat output and quality by identifying wheat diseases quickly and accurately, an Android phone-based system for intelligently diagnosing wheat diseases is proposed. The users collect images of wheat diseases using Android phones and send the images across the network to the server for disease diagnosis. After receiving disease images, the server performs image segmentation by converting the images from RGB color space to HSI color space. The color moment and the gray level co-occurrence matrix are utilized to extract the color and texture features of the diseases. The preferred features are input to the support vector machine for recognition. And the identification results are fed back to the client. Application tests demonstrate the ability of the proposed system to identify types of disease accurately and efficiently. Thus, the proposed system is very feasible and has a bright prospect for widespread adoption. Keywords: Wheat Disease; Android System; Disease Diagnosis; Support Vector Machine 1 Introduction Due to lack of the ability of disease diagnosis when the crops are infected, farmers are difficult to make reasonable and effective prevention and treatment. Once adopting improper measures, it might lead the serious reduction of crops. Uncontrolled use of pesticide not only lowers the quality of production, but also harms human health [1, 2]. Therefore, providing guidance for farmers on how to identify diseases effectively is a key issue in agriculture production. The former disease diagnoses were generally based on experience or identification of pathogen in the lab, which caused wrong judgments or missed the disease control period because of the long time [3, 4]. In recent years, market share of Android smart phones keeps rising [5]. Free and opensource Android operating system provides flexible independent design space for software designers, thus providing a good development platform for developing an open-source, free and easy-touse agricultural intelligent system. Currently, the diagnosis of crop diseases largely depends on Project supported by the National Nature Science Foundation of China (No. 61302118). Corresponding author. Email address: xyqmouse@163.com (Yongquan Xia). 1548 7741 / Copyright 2015 Binary Information Press DOI: 10.12733/jics20150084

6846 Y. Xia et al. / Journal of Information & Computational Science 12:18 (2015) 6845 6852 the expert supporting system, which takes the textual description of plant diseases and insect pests as the reasoning basis. Due to their limit in expertise, the farmers cannot describe the features of plant diseases accurately and professionally. Hence, the expert supporting system is not very effective in practice. To conduct remote diagnosis over crops by effectively combining smart mobile with image processing technology can enable farmers to timely and accurately adopt corresponding measures to prevent and treat diseases and reduce losses, which has great significance for high production and optimizing production of crops. In this paper, a design method of intelligent disease diagnosis system based on Android platform is proposed. The system uses Android phones to collect disease images, simplifies image background by man-machine interactive mode, and adopts the methods of mobile wireless communication to send images to the server. The server segments disease images, computes disease feature values, uses support vector machine to identify and sends identification result to the Android phone. The system has many advantages such as simple operation, fast diagnosis speed and accurate diagnosis result, and can effectively diagnose wheat diseases in farms 2 Overall System Designs 2.1 Development Environment The system development environment is JDK 7 (Java Development Kit) + Eclipse 4.4.2 + Android SDK (Software Development Kit) + ADT (Android Development Tools) + Tomcat 7.0. JDK provides java compilation environment; Eclipse is an expandable development platform based on Java; Android SDK provides Android application development environment and instrument; ADT is a universal plug-in that supports Android project, making the creation, operation and debugging of Android application faster and more convenient; Tomcat is a small lightweight application server. 2.2 Steps for Plant Disease Diagnosis Basic steps for wheat disease diagnosis: (1) Collect disease images. Use the Android phones to collect disease images and crop the images via human-computer interaction. (2) Transmit disease images. Send the cropped images to the server through the http protocol. (3) Segment the disease images. Convert RGB color space to HSI color space and segment images in HSI color space. (4) Extract disease features. The selected seven feature variables are used as the input vector for disease recognition. (5) Diagnose the disease images. Use the RBF-based SVM as the recognition model to diagnose the diseases, and send the diagnosis results to the Android phones.

Y. Xia et al. / Journal of Information & Computational Science 12:18 (2015) 6845 6852 6847 3 Segmentation of Disease Images Accurate segmentation of disease images is a prerequisite for disease recognition [6]. Tailored sub-images only contain leaf healthy areas and disease areas. The green pixels mainly indicate healthy areas of leaves without containing any valuable disease information while disease areas generally are not green. According to the features of disease images, disease areas are obtained after eliminating green pixels. The correlation among three primary colors of red, green and blue in RGB color model is very strong, so we can not use a single parameter to designate a range for green pixels in the image. Therefore, we can not conduct threshold value segmentation over the image by relying on setting r, g and b. In HSI color space, hue, saturation and intensity are independently decoupled and can accurately quantify color information of pixel points [7]. Therefore, the RGB color space is switched into HSI color space to set hue scope to eliminate green pixels. Additionally, the larger the saturation becomes, the closer to being pure the color gets, while the smaller the saturation becomes, the closer to being grey the color gets. Intensity does not affect hue significantly, so it needs not to be limited anymore. The paper conducts plenty of sampling over different parts of hue on the image to generate the formula of eliminating leaf healthy areas: 60 < H < 180 0.177 < S 1 (1) Wheat blades suffering powdery mildew, stripe rust, leaf rust and stem rust are segmented by the use of the OTSU method, the iteration method and the proposed method, respectively. The segmentation results are shown in Fig. 1. (a) Original images (b) OTSU method (c) Iteration method (d) Proposed method Fig. 1: Segmentation results

6848 Y. Xia et al. / Journal of Information & Computational Science 12:18 (2015) 6845 6852 From the above experimental results, it can be seen that the grey levels of blades suffering powdery mildew, stripe rust and leaf rust are all less than that of healthy blade areas, while the grey level of blades suffering stem rust is larger than that of healthy blade areas. Although the disease areas and the healthy areas are segmented into two parts, the computer is unable to determine the segmented disease areas, thus impeding automatic recognition of diseases. The proposed algorithm in this paper can not only segment effectively, but also distinguish between healthy and disease areas by checking whether the pixels are green. This enables the computer to accurately calculate the feature values of disease areas and recognize the diseases. Therefore, the proposed method is chosen to segment blade disease images during disease recognition. 4 Extraction of Disease Features 4.1 Texture and Color Features Texture feature refers to the change of image gray level, which is related to spatial statistics. Gray level co-occurrence matrix, proposed by Haralick, is an important method to analyze the image texture features, which can accurately reflect the texture roughness and repeated direction [8]. Haralick defined fourteen gray level co-occurrence matrix parameters for texture analysis. Ulaby found that only four features are not correlated within fourteen texture features based on gray level co-occurrence matrix [9]. It is easy to count four features with higher classification accuracy. The following four parameters are applied to extract the texture features of wheat diseases. The formulas are as follows: Energy = [P (i, j, d, θ)] 2 (2) Contrast = Entropy = Correlation = i=1 i=1 i=1 i=1 (i j) 2 P (i, j, d, θ) (3) P (i, j, d, θ) log P (i, j, d, θ) (4) (i µ i ) (j µ j ) P (i, j, d, θ) σ i σ j (5) In Eq. (2) to (5), P (i, j, d, θ) means gray level co-occurrence matrix, L means the width of gray level co-occurrence matrix, while µ i, µ j separately refers to the mean value of P (i) and P (j) and σ i, σ j represent variance of P (i) and P (j). Stable color feature is an important parameter during the disease identification process. Color moment is a simple and effective color feature representation method proposed by Stricker and Orengo [10]. There are first moment, second moment and third moment, etc. The color information is mainly distributed in the lower order moment, so first moment and second moment are chosen to be the color feature. The formulas are as follows: M i = 1 N N P ij (6)

Y. Xia et al. / Journal of Information & Computational Science 12:18 (2015) 6845 6852 6849 [ 1 V i = N N (P ij M i ) 2 ] 1 2 (7) In Eq. (6) and (7), P ij means the gray level of j pixel in the i component, while N refers to the sum of pixels. 4.2 Feature Selection Different feature variables make different contributions to plant disease recognition, so selecting proper feature variables is needed. Mean Impact Value (MIV) reflects the changes of the neural network weight matrix, which is considered as one of the best index to evaluate variable correlation. Compared with the traditional artificial neural network, the support vector machine has a simpler structure and substantially increases generalization ability. The method of the feature variables selection is based on the mean impact value and support vector regression is presented. The algorithm process is as follows: (1) The training sample S is used to train the support vector machine. (2) 10% is added and subtracted from each independent variable in the training sample S, and then get two new samples: S1 and S2. (3) Do regression prediction for S1 and S2, and get two results: M1 and M2. (4) The difference value of M1 and M2 is the impact value W based on the output after the changes of corresponding independent variables. (5) While W is averaged based on the number of observation cases, the mean impact value MIV of independent variable on the dependent variable is obtained. Similarly, we can obtain MIV for each independent variable. (6) Sort the variables according to their sizes of absolute value of MIV, and then get sequence table, thus realizing the variable selection. The simulation experiments are conducted in Matlab 7.0, eighty images in total (twenty for each) are selected as training samples in term of wheat powdery mildew, stripe rust, leaf rust and stem rust. The feature variables are sorted based on effect degree of recognition results. The sorting results are shown in Table 1. Table 1: MIV sequence Rank Feature variables MIV Rank Feature variables MIV 1 first moment of G component 0.0264 8 mean value of contrast 0.0034 2 second moment of G component 0.0144 9 standard deviation of energy 0.0032 3 second moment of R component 0.0075 10 standard deviation of correlation 0.0024 4 second moment of B component 0.0068 11 standard deviation of entropy 0.0015 5 first moment of R component 0.0067 12 mean value of energy 0.0004 6 mean value of entropy 0.0053 13 standard deviation of contrast 0.0003 7 first moment of B component 0.0043 14 mean value of correlation 0.0001 According to the MIV sequence, the first feature variable is used as the input vector to train the support vector machine and obtain the first prediction model. Then, the first two feature

6850 Y. Xia et al. / Journal of Information & Computational Science 12:18 (2015) 6845 6852 variables are used as the input vector to train the support vector machine and obtain the second model. Similarly, the first fourteen feature variables are used as the input vector to obtain the fourteenth prediction model. The model accuracy is evaluated via ten-fold cross-validation [11]. The classification accuracy of the fourteen prediction models is shown in Fig. 2. 100 95 Accuracy/% 90 85 80 75 Ten-fold cross-validation 70 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Number of feature variables Fig. 2: Accuracy of prediction models From the above experimental results, it can be seen that the classification model will learn too many features during the training and the inherent laws of samples cannot be represented by redundant feature variables, resulting in low prediction performance of the prediction model. The accuracy of prediction model is the highest when seven feature variables which have the greatest impact to classification result are used to identify powdery mildew, stripe rust, leaf rust and stem rust. 5 Disease Recognition In actual production, farmers conduct judgment mainly by relying on their own experiences and perceptions, which seriously affect the accuracy of disease forecast. Moreover, farmers usually live far away from the experts, which makes it time-consuming and inconvenient to consult them [12]. To conduct diagnosis over diseases by adopting support vector machine can timely, rapidly and accurately generate the result. Support vector machine is a machine study method proposed by Vapnik in 1995 based on VC dimension theory of statistical theory and minimum risk principle of structure to seek best compromise in complexity of model and studying ability in accordance with limited sample information, with the purpose of obtaining best promotion ability [13]. Some interfaces of the system are shown in Fig. 3. 6 Software Test Software test is an important component of software development. Programs are packaged into Android installation package files, which are installed into MI 1S mobile whose system version is Android 4.0. Twenty images are selected for each of the disease (i.e., powdery mildew, stripe

Y. Xia et al. / Journal of Information & Computational Science 12:18 (2015) 6845 6852 6851 rust, leaf rust and stem rust) to do the test. According to the test, the diagnosis times in WIFI environment and GPRS environment are slightly different, affected by the sizes of images and network qualities, which lie between five and ten seconds. The diagnosis results are shown in Table 2. (a) Diagnosis way interface (b) Tailoring interface Fig. 3: System interfaces (c) Diagnosis result interface Table 2: Diagnosis results Disease name Correct number identified (total) Accuracy/% Powdery mildew 20(20) 100 Stripe rust 18(20) 90 Leaf rust 19(20) 95 Stem rust 19(20) 95 Total 76(80) 95 7 Conclusion The paper proposes a design method of wheat disease diagnosis system based on Android platform, and the system is not subject to factors such as time, location and distance, with simple and convenient operation. It reduces misdiagnosis and solves the issue that farmers cannot consult experts timely, which effectively improves output and quality of wheat.

6852 Y. Xia et al. / Journal of Information & Computational Science 12:18 (2015) 6845 6852 References [1] J. D. Pujari, R. Yakkundimath, A. S. Byadgi, Classification of fungal disease symptoms affected on cereals using color texture features [J], International Journal of Signal Processing, 6(6), 2013, 321-330 [2] J. G. Dai, J. C. Lai, Image-rule-based diagnostic expert system for cotton diseases and pests based on mobile terminal with android system [J], Transactions of the Chinese Society for Agricultural Machinery, 46(1), 2015, 35-44 [3] Q. Yang, X. Y. Gao, J. L. Wu, et al., Identification of barley diseases based on texture color feature [J], Journal of China Agricultural University, 18(5), 2013, 129-135 [4] S. Arivazhagan, R. N. Shebiah, S. Ananthi et al., Detection of unhealthy region of plant leaves and classification of plant leaf diseases using texture features [J], Agric. Eng. Int.: CIGR Journal, 15(1), 2013, 211-217 [5] L. N. Yang, L. T. Gao, E. S. Lin et al., Intelligent diagnose system of diseases and insect pests in sweet corn based on mobile terminal with android system [J], Transactions of the Chinese Society of Agricultural Engineering, 28(18), 2012, 163-168 [6] R. Ma, Y. P. Cheng, Y. H. Gu, An approach of removing stroke based on touching pattern for handwritten digits egmentation [C], 2013 International Conference on Advanced Computer Science and Electronics Information, 2013, 237-240 [7] Y. M. Xie, Z. Fan, S. L. Zhang, Color image edge detection based on HSI color space [J], Computer Engineering, 39(9), 2013, 12-19 [8] R. M. Haralick, K. Shanmugam, I. Dinstein, Texture features for image classification [J], IEEE Transactions on Systems, Man and Cybernetics, 3(6), 1973, 610-621 [9] F. T. Ulaby, F. Kouyate, B. Brisco et al., Textural information in SAR images [C], IEEE Transactions on Geoscience and Remote Sensing, 24(2), 1986, 235-245 [10] M. Stricker, M. Orengo, Similarity of color images [J], SPIE Storage and Retrieval for Image and Video Databases III, 2(4), 1995, 381-392 [11] N. Yuan, P. Yang, Z. J. Liu, Gait recognition by the mean impact value and probability neural network [J], Journal of Harbin Engineering University, 36(2), 2015, 181-185 [12] Q. Z. Zhao, G. C. Jin, W. J. Zhou et al., Information collection system for diseases and pests in cotton field based on mobile GIS [J], Transactions of the Chinese Society of Agricultural Engineering, 31(4), 2015, 183-190 [13] C. Cortes, V. Vapnik, Support vector networks [J], Machine Learning, 20(3), 1995, 273-297