Recognition And Early Warning For Major Vegetable Pests In South China Based On Machine Learning

With warm weather,abundant rainwater and fertile soils,South China is suitable for vegetable growth.Guangdong Province ranks first among cultivated areas in production output and consumption amount of vegetables countrywide.However,pests are inevitable during cultivation.Notably,the moist weather in South China is highly suitable for proliferation of pests.Pests reduce output and quality of vegetables.Agricultural personnel often spray pesticides for pests control,resulting in excessive pesticide residues in vegetables,causing serious harm to human health.Furthermore,excessive application of pesticide leads to severe environmental pollution.To effectively prevent pests infestation and reduce use of pesticides,it is necessary to apply modern advanced information technologies to the pests control of vegetables.Pests will be detected by establishing a vegetable pest monitoring and alarm system based on the intelligent cloud platform.In this way,pests can be found early,the types and quantities of pests can be understood,and corresponding alarm information can be provided to achieve timely control of small-scale pests.This will reduce the use of pesticides and economic losses caused by pests,improve the level of integrated pest management,achieve precise control of agricultural pests,and promote the development of precision agriculture.With the advancement of modern information technology,the application of intelligent technologies such as machine learning has become wider,and has good application effects in various industries.Computer vision technologies and machine learning have seldom been applied to the research and supervision of dynamic regulations of vegetable pests so far.Aiming at the serious situation of vegetable pests in South China,the investigation and research on the main pests such as whiteflies,striped flea beetle,diamondback moths and thrips are carried out,and a serious pests monitoring and identification alarm system is established.The thesis mainly includes the following works:(1)In this paper,a pests image data sampling node device suitable for scene illumination changes is designed.Traditionally,the number and type of pests are manually collected in the field,which is time-consuming and labor-intensive.The pests image data acquisition node device designed in this paper can be used to automatically collect pests image data from a long distance.The node device includes a pests device with an pests trap module,a supervisory module that can collect images of pests,and a solar module that can provide power to the device.(2)This paper proposes a vegetable pest identification method based on morphology learning and machine learning.Firstly,the region is identified by the trapping plate based on the HSV color space.Image denoising and hole filling are proceeded based on the morphological method.Secondly,the random forest edge detection algorithm is used to separate and extract targets.Third,the SIFT is used to extract features of the target pests image block.K-means are used to cluster the characteristics of pests.Fourthly,the bag of feature(BOF)model is used to characterize the pests images,construct the BOF model of vegetable pests,and establish a visual dictionary of pests.The dictionary is then used to process pests images,to count visual vocabulary frequencies in pests image blocks,and to construct descriptive vectors of pests images.Fifth,the BOF model and the SVM classifier are combined for classification,identification and enumeration of vegetable pests.(3)The regression alarming model for vegetable pests occurrence is established.Firstly,clustering grading is conducted to vegetable pests quantity acquired from machine learning with K-means.Secondly,environmental data collected by environmental sensors is treated,which is then conducted with dimension reduction with the principal component analysis,to acquire the principal component factors.Finally,the regression model is utilized to establish the alarming model related to vegetable pests occurrence grade and environmental factors.The level of vegetable pests is acquired based on parameters of environmental factors to provide assistance in accurate guidance for agricultural production.(4)The serious pests supervision and recognition alarming platform in South China is established.The platform can be utilized to collect images of vegetable pests automatically.In addition,it conducts pests recognition and counting from acquired images,and reflects quantity information of pests to vegetable cultivation personnel rapidly.It has wide application prospects.(5)Vegetable pests recognition and supervision experiments and related analysis are conducted.Meteorological sensors,soil sensors and wireless network nodes are deployed in farms and scientific research bases in 3 different administrative regions in Guangzhou.To acquire high-resolution pests images and environmental and meteorological data,each region deploys as a similar experimental scheme.In this way,it is possible to construct the vegetable pests alarming model between environmental meteorological factors and pests quantity,and develop cross-regional comprehensive network tests remotely.Based on experimental analysis on network data transmission speed of pests data collection nodes,the average data transmission speed is 166.77 kbps.Taking energy consumption conditions and node transmission data packet loss probability into consideration,the size of the transmission data package is determined to be within 1KB.The recognition rate of the vegetable pests is analyzed based on algorithm counting and artificial counting to vegetable pests images.The result is 91%,meeting the needs of practical application.Finally,correlation analysis between pests quantity and environmental meteorological data is conducted,to establish the vegetable pests supervision alarming model.The prediction accuracy achieves 84%,which can provide support to pests prediction and alarming for vegetable cultivation personnel.The main innovation points of the thesis are: A wireless collection node for pests supervision adaptive to scene illumination variation is designed.A machine learning based classification and identification algorithm is proposed,to improve the robustness of detection by combination of various algorithms such as mathematical morphology,K-means and visual dictionary.It is adaptive to some regional shape variations and illumination changes during matching.Four main typical pests,namely whiteflies,striped flea beetle,diamondback moths and thrips are identified through images.It constructs the vegetable pests alarming model between environmental meteorological factors and quantity of pests.In addition,it provides important technical and information supports for development of precision agriculture,and provides support for the decision-making of agricultural production units and management units.In turn it provides guidance for relevant personnel to timely adopt preventive measures in countering infestation.