Font Size: a A A

Research On Rice Blast Detection Techniques Based On Multispectral Vision

Posted on:2010-01-28Degree:DoctorType:Dissertation
Country:ChinaCandidate:L QiFull Text:PDF
GTID:1118360272996808Subject:Agricultural Electrification and Automation
Abstract/Summary:PDF Full Text Request
Rice is the most important food crop in China, and rice production is also responsible for food security. However, pests and diseases that have bad impact on rice yield and quality invade rice usually. Rice blast is one of the most serious diseases in Southern and Eastern plant area in China.Results of naked-eye observing rice blast are different because of subjective factors, and the method of observations is laborious and time consuming. Although results of detecting rice blast are precise based on molecular biology technique, this kind of method cannot apply to real time field operations. A new technique is in urgent need for detecting rice blast fast and accurately because of above reasons.This study summarized multispectral analysis technique, image processing technique and multispectral vision technique applied in diseases and insect pests detection technique at home and abroad. Starting with the theory of green plant disease detection techniques based on multispectral vision, the study focused on rice seedling and canopy infected by rice blast as research object. According to the prospective mission requirement of equipment research center of national rice industry technology system and the goal of sub-project"Variable automatic control fertilizer and diction mechanism development"(project number:2006AA10A305-7) of National High-tech R&D Program (863 Program), the study comprehensively adopted advanced thoughts on the area of Modern Agricultural Engineering Technology,Electronic Information Technology(Computer Vision, Virtual) and Artificial Intelligent Technology (Artificial Neural Network, Support Vector Machine, Time Series Analysis) to research key technologies on rice blast disease gradation and early detection.The main contents and results are listed as follows. According to research demand, Seedlings samples needed by experiment were prepared, and hardware platform of multispectral vision system was designed. The samples preparation included cultivation of magnaporthe grisea, cultivation of rice seedlings and Inoculation. Multispectral vision system hardware platform consisted of multispectral image acquisition system, control system and illumination system. The image acquisition system was composed of multispectral camera (MS3100) which was capable of acquiring images in 3 wave bands (NIR, R and G) and an image acquisition device (NI1428). Control system consisted of portable computer and camera control software (DTControl). The main body of illumination system was self-designed light box which was two-layer structured. In the upper layer, adjustable mechanism was designed so that shooting height and scope of the camera could be adjusted flexibly. In the lower layer, halogen lamps covered by PVC panel was arranged around the box, providing the system stable light condition.Virtual instrument technique and multispectral vision technique were combined firstly in this study to develop function-module-extensible multispectral vision software. Labview and its vision module were used to design multispectral vision processing software which consisted of multispectral image acquisition, storage and display module and image real time processing and analysis module on the multispectral platform, so that rice seedling automatic recognition and image information statistics could be achieved in either lab condition or field condition. Concerning image segmentation algorithm, the difference vegetation index image was selected as treat image calculated by subtraction of image gray value in near infrared band and red band. Compared with the original images, the difference vegetation index image was easy to use and highlighted the gray value difference between rice seedling and the background, so that the system could precisely and timely recognize rice seedling using global automatic threshold method. In lab conditions, multispectral images of 500 potted rice seedling were acquired and processed, rice seedling was recognized precisely and over 12 kinds of multispectral image information(gray average in infrared, red and green wave band, gray standard deviation, rice seedling area, ratio vegetation index, difference vegetation index and normalized difference vegetation index) of each group were obtained and recorded. In field conditions, 212 multispectral images were acquired and 210 were correctly segmented, with segmentation accuracy of 99%. Under window of 2.4m×1.8m, average processing time of each image group was 311ms, meeting the demand of processing in real time.Probabilistic Neural Network and Support Vector Machine were applied in rice blast grading detection. Firstly, disease gradation was defined as five according to chemical application and International Rice Institute leaf blast resistance grading standard. Then, through the use of multispectral vision recognition technique, multispectral image in red band was obtained as segmentation target, and blast spot was recognized for infection degree (spot area/seedling area) statistics so that disease grade could be determined in accordance with grading standard defined in this research. After correlation analysis, gray value in red and green band images and ratio vegetation index whose correlation coefficients with infection degree being 0.947, 0.882 and 0.941 respectively were chosen as characteristic parameters for grading detection model. Finally, grading detection model was established by PNN and SVM technology, with 150 chosen samples as research objects. When PNN was used to establish model and radial basis function distribution density chosen as 0.5, 92 among 102 testing samples were correctly graded, with gradation accuracy being 91.18%. When SVM was used to establish model, LibSVM embedded grid-search parameter optimization model was used, and penalty factor C of support vector machine determined as 2048, kernel parameter gamma as 2, 98 among 102 samples were correctly graded, with gradation accuracy being 96.08%. Analysis on samples wrongly graded showed physiological litter, chose of borderline samples and classification algorithm were three key reasons of classification errors. The results showed that accuracy of model trained by SVM based on the structural risk minimization principle was better than that of PNN whose principle was empirical risk minimization.Multispectral vision technique and time series technique were cooperated firstly to study early detection of rice blast. The gray value of NIR image was selected as characteristic parameter for early detection of rice blast by analyzing significant differences on mean value of samples'characteristics between inoculation and no-inoculation samples and comparing change trend of seedling and canopy multispectral image information. The time series of NIR image gray value was set up, and then the similarity of different time series was computed by Euclidean distance and pattern distance respectively. The early classification detection of rice blast was accomplished by K-nearest neighbor arithmetic. The classification accuracy of infected seedlings samples'early detection is 70.6% using Euclidean distance, and 94.9% using pattern distance. The results showed that pattern distance is more suitable for describing the change trend of time series.In a word, key techniques of rice blast detection researched in the study provided theoretical foundation and technique support for variable sprayer and disease forecast, and broke a new path in plant disease detection.
Keywords/Search Tags:Agricultural engineering, Multispectral vision, Rice blast detection, Virtual instrument, Probabilistic neural network, Support vector machine, Time series analysis
PDF Full Text Request
Related items