Identification Of Invasive Plant Mikania Micrantha Based On Hyperspectral Analysis

As a "plant killer",Mikania micrantha is now one of the 100 most invasive alien species in the world.It has spread rapidly in Guangdong,Guangxi,Hainan and Yunnan in recent years,seriously damaging the ecological environment.It is necessary to control the diffusion trend of mikania micrantha through intelligent monitoring and early prevention.At present,the image recognition method and hyperspectral monitoring method are often used in the monitoring of alien invasive plants.In particular,the hyperspectral monitoring method is more widely used because of its features of high spectral resolution,multiple bands and abundant information.Therefore,researches related to the application of hyperspectral technology to the intelligent monitoring of mikania micrantha should be carried out,and the key to the intelligent monitoring of mikania micrantha lies in whether it can detect and identify mikania micrantha targets from the complex environment in the field.In this paper,mikania micrantha was taken as the research object,and the recognition method of mikania micrantha in the field was studied by combining hyperspectral dimensionality reduction technology and pattern recognition technology based on support vector machine.The research contents are as follows:(1)Hyperspectral data collection and sample acquisition.Hyperspectral data and images of invasive plant mikania micrantha were collected by non-imaging hyperspectral acquisition system in laboratory and hyperspectral image acquisition system in field,and hyperspectral data samples based on the two data sources were obtained.(2)Study on the method of hyperspectral data preprocessing.Due to environmental interference or human factors,not only some data anomalies or abnormal data loss,but also sample data noise,baseline drift and other issues may be caused.Abnormal sample data removal and spectral data preprocessing were performed on mikania micranthus hyperspectral data samples from the two data sources.As for the pretreatment methods,the laboratory mikania micranthus hyperspectral data were smoothed and then differentiated,and the field hyperspectral data were pretreated with eight methods respectively.Eight preprocessing methods are moving average smoothing(9 points),S-G smoothing,first order differential,second order differential,multiple scattering correction,standard normal variable transformation,first derivative and then smoothing(first order differential+S-G smoothing),first standard normal variable transformation and then trending processing(SNV+Detrending).(3)Study on dimensionality reduction method of hyperspectral data.In order to solve the problem of large-band hyperspectral data,large information redundancy and excessive calculation of mikania micrantha based on full-band,two different dimensionality reduction methods are proposed,which are mainly used in the field of Mikania micrantha spectral data after dimension reduction pretreatment.They are full-band feature extraction of hyperspectral data based on principal component analysis method and spectral feature selection of mikania micrantha based on local band.Feature extraction is the principal component analysis of hyperspectral data on the basis of nine different pretreatment methods(no pretreatment and eight pretreatment methods).The spectral feature selection of mikania micrantha is to select local characteristic bands based on the analysis of hyperspectral data curve characteristics of laboratory mikania micrantha.(4)Establishment of classification model based on support vector machine and research of feature set optimization method.Support vector machine models are established based on two dimensionality reduction methods.The results showed that the classification model corresponding to SNV-PCA feature set had the highest recognition accuracy of 92.96%for mikania micranthus among the nine classification models established after dimensionality reduction by principal component analysis,and the classification model corresponding to feature set selected based on spectral features of mikania micranthus had the highest recognition accuracy of 90.14%.The two feature sets were further optimized based on single feature ranking method.The feature set obtained from the spectral feature selection of mikania micrantha has a better optimization effect,and its corresponding recognition accuracy is 90.85%,which improves the recognition accuracy and model stability.(5)Test verification and result analysis.The prediction set hyperspectral data was input into the established SVM classification model by SNV-PCA feature set and optimized feature set based on mikania micrantha spectral features,and the recognition accuracy was 81%and 84%,respectively.Therefore,it is believed that the SVM model established by feature extraction method based on principal component analysis can be used for the identification of invasive plant mikania micrantha,and the identification result is optimal when the dimension of SNV-PCA feature set is 10;The SVM model based on the method of selecting the spectral characteristics of mikania micrantha is also suitable for the identification of mikania micrantha.Among them,518nm,538nm,550nm and 710nm are the sensitive bands for target identification of mikania micrantha,and 490-530nm,550-582nm and 640-760nm are the sensitive bands for target identification of mikania micrantha.