Classification of sunn pests using sound processing methods and an embedded system realization

Cereal types are the plant types that are widely planted and consumed as food all around the world. With their weight in different domestic cuisines, production of cereals holds an undeniably great economic importance. This importance places great demand on the detection and the struggle against or control of insect pests of cereals. Wheat is one of the types of cereals. Consumption of wheat based products covers major share in the nutrition habits of inhabitants of Republic of Turkey. As a direct consequence, vast areas are designated as wheat plantations in Turkish agriculture. With respect to importance of wheat by means of production and consumption, it is demanding that product yield of wheat should be reliable, healthy and abundant by years. Apart from damages caused by unexpected seasonal changes, the efficient control of pests of wheats has vital significance. Sunn pests, such as Eurygaster and Aelia species are pest insects that have been clarified as harmful to Turkish wheat production for years. Although having difference in population density, having similar seasonal behaviors, the sunn pests can be encountered nearly all parts of Turkey. Sunn pests are the most important group of insect species which damage cereal plantations. Proposing early detection and control procedures of these insects would be significantly beneficial for planning and reliability of agricultural activities. In this thesis, a system model with the purpose of early detection and classification of sunn pests are proposed. An embedded system based on the system model is given as an exemplary application. First chapter of the thesis is designed as an introduction chapter and the purpose of the work, which is presented above is extended in this section. The hypothesis can be summarized as "sunn pest can be classified with high accuracy with machine learning methods by using sound feature vectors obtained from the data with high sampling frequency and this method can be successfully applied to early detection applications." This chapter contains, literature review, discussion of needs for experiments and information about data collection procedure. This work carries an interdisciplinary character because of its nature as it contains insect biology and electronics engineering. Therefore, information about entomology and plant protection is presented. Moreover examples of classification of insect species, the classification methods, sensors and setups that are used in insect classification tasks and usage of sound data of insects in insect classification procedures are described. In this thesis four species of sunn pests are examined. The species are members of Eurygaster and Aelia genus. General definitions and information about four insect classes of Eurygaster Integriceps, Eurygaster Austriaca, Aelia Rostrata and Aelia Accumniata are given in the first chapter. The sound recording procedure, the hardware and software used in this recording and the recording system produced for sound recordings of subject insects are introduced in this section too. Second chapter of thesis report is sorted for mathematical background which the thesis is based on. In the thesis, especially well known and widely applied sound processing, analysis and automatic speech recognition methods are preferred. The mathematical equations are used for introduction of respective methods that are used for analyzing sound data. For preliminary analysis, Zero Crossing Rate, Energy, Entropy of Energy, Spectral Centroid, Spectral Spread and Spectral Entropy information are gathered for midterm statistics of sound data. These methods give general information for subject sound data but, the analysis data are not preferred for classification task. For classification task, well known speech feature extraction methods of Linear Predictive Cepstral Coefficients, Mel Frequency Cepstral Coefficients and Line Spectral Frequencies methods are preferred and applied. Continuing the chapter the machine learning approach for classification is introduced. The machine learning method that is used for the automatic classification of sound feature data is chosen as simple k-Nearest Neighborhood method. Third chapter is the section of thesis, where the system model which is produced after the application of mathematical methods that are introduced in the chapter two is described. Construction of system model and the performance analysis methods related to this model are examined in this section. The performance metrics are defined with respect to Bayesian Error Rates and Confusion Matrices. After the introduction performance analysis metrics, the chapter concludes with the presentation of intermediary performance analysis results of system application on MATLAB software environment. These intermediary results are composed of data such as histograms from obtained midterm statistics, Bayesian Error tables, confusion matrices, system accuracy and precision tables and neighborhood parameter-system accuracy graph. The data that are used in the machine learning part of system model are constructed with the combinations of feature vectors that are obtained with three feature extraction method which are explained in Chapter Two. The input vectors for k-Nearest Neighborhood procedure is obtained for each recording using the combinations of Linear Predictive Cepstral Coefficients, Mel Frequency Cepstral Coefficients and Line Spectral Frequencies methods outputs. The behavior of system performance is examined for each combination and through this examination the most successful combination is decided as the usage of Mel Frequency Cepstral Coefficients and Line Spectral Frequencies together. Fourth part of the thesis is about embedded system approach. In this chapter, a setup that depends on the Mel Frequency Cepstral Coefficients and Line Spectral Frequencies combination, which was decided as the most successful result, is applied on the UDOO QUAD microcomputer. Being open source and enabling applications of numerical method libraries, Python programming language is preferred during the application of setup. Encountered problems about reading sound data on microcomputer and the solution to this problem is also explained in this chapter too. The cause of re-establishing system model this microcomputer is to show that the works beginning with this thesis work are capable of an integrated design which can be applied to on-field problems. The fifth part of the thesis report contains conclusion and feature works. In conclusion of thesis report includes revaluation of intermediary results, in addition to confusion matrix, accuracy and precision results obtained from embedded system implementation. As the evaluation for embedded system application shows clearly that the a design for on-field application is subject to revaluation of programming language, application algorithm and the microcontroller that would be used for application, the established mathematical model and applied setup make it possible to say that hypothesis was verified. The proposed method in the thesis study is a first study in the direction of aimed purpose. This first step concludes its duty as reaching aimed parameters so that it provides a solid ground for feature works for cheaper, more efficient and more precise commercial or scientific applications. The last chapter ends with this evaluation and discussion about aforementioned future works.