Research On Classification Strategy Of Agricultural Plants Based On Leaf Spectrum

The traditional method of vegetation resource survey is time-consuming and difficult to be updated;moreover,multi-spectral remote sensing data tends to experience the phenomenon of different objects with same spectrum,which can only distinguish the objects(e.g.vegetation,water and bare land)with widely different reflectance.However,the multi-spectral remote sensing images cannot meet the classification requirements at the species level.In contrast,hyperspectral remote sensing data has significant advantages such as high spectral resolution and a large number of bands,and it provides more detailed spectral information of the ground objects,thereby vegetation can be classified more precisely and accurately.Previous studies have rarely compared different types of spectral transformation,spectral feature variables,and machine learning algorithms to classify vegetation in agricultural regions,and further it is difficult to recommend reasonable method combinations or classification strategies for researchers.With laboratory-based leaf hyperspectral reflectance of leaves collected from eight types of vegetation in agriculture regions of Yixing,Jiangsu,China,this thesis explored the optimal strategy of vegetation classification.First,three types of spectral feature variables(three-side parameters,absorption feature parameters,and vegetation indices)were extracted using first derivative and continuum-removal methods based on the preprocessed reflectance.Secondly,discrete wavelet analysis was employed to decompose the original,first derivative and continuum-removed spectra into approximate and detail spectra.In order to avoid the dimension disasters caused by the use of hyperspectral data,the principal component analysis method was employed for dimension reduction with the original,first derivative and continuum-removed spectra as well as the detailed spectra at the first to eighth decomposition level.Finally,the single spectral feature variables,full spectra(original,first derivative and continuum-removed),and the first 10 principal components of the full spectra and their corresponding detail spectra were taken as the independent classification variable,three kinds of machine learning algorithms(k-nearest neighbor,support vector machine,and random forest)were used to classify 8 types of vegetation,and classification accuracy under different classification strategies were calculated and evaluated.The main research results of this thesis are as follows:(1)The three methods with single spectral feature variables achieved low overall classification accuracy less than 50%.With the principal component analysis of the original,first derivative and continuum-removed spectra,the overall classification accuracies was greatly improved,in which the classification model with the first derivative reflectance performed best with the overall classification accuracy nearly 95%.(2)Compared with the k-nearest neighbor and random forest methods,the support vector machine method obtained better classification accuracies.The SVM model with the principal component of the original,first derivative and continuum-removed spectra obtained average accuracies over 90%.(3)The classification models with the combination of principal component analysis and discrete wavelet transform outperformed the models with merely principal component analysis,and the classification model performed best with the detail spectra at the fifth or sixth decomposition level.(4)Among the four wavelet functions(Db5,Db3,Haar and Coif5),the Haar wavelet obtained the best classification results.The three classification models with the detail reflectance spectra at the first to eighth decomposition level achieved the average classification accuracy of 76.8 and Kappa coefficient over 0.7.With the analysis of the leaf spectral features,the vegetation classification results under different classification strategies were evaluated and compared.The results indicated that the principal component analysis combined with the support vector machine method can meet the accuracy requirements of the vegetation classification,and the combination of principal component analysis and discrete wavelet decomposition is recommended for vegetation classification.The results in this thesis can provide technical references and theoretical support for vegetation remote sensing classification,precision agriculture,and resource investigation at the landscape or regional scale.