| The classification of plants and monitoring of their growing status in a certain area is of great significance for agriculture,ecology,urban green land management.The traditional way to conduct the plants classification and growing status monitoring are mainly based on manual investigations,which are time-consuming,laborious and inefficient.Hyperspectral remote sensing data has great potential for classification of plants and monitoring of their growing status due to its rich spectral information and the ability to conduct large-area vegetation monitoring.The spectral analysis technique for classification of plants and monitoring of their growing status is a basis for constructing a plant spectral library,which can support large-scale plant monitoring.At present,a variety of spectral libraries have been established for different purposes,but the spectral library specific for vegetation monitoring is rare.Besides,a systematic study on the corresponding features selection and classification algorithms are needed.To deal with these problems,this paper conducted severalexperiments and collect data that can support the study on spectral feature optimization and development of classification algorithms.Besides,a feature selection procedure that consisted by feature sensitivity ananlysis and feature robustness analysis are proposed.Several machine learning algorithms were used and compared in classification of plants and monitoring of their growing status.In addition,an UAV hyperspectral image data was collected to test corresponding methods.A prototype spectral library for vegetation monitoring was then established,which included several important functions.The main contents in this paper can be concluded as:(1)Conducted systematic experiments to obtain data for research.A number of experiments were carryied out,which collected a set of plant spectral data and corresponding ground survey data.A total of 1910 standard spectra were collected from 77 plants in 3 locations in Hangzhou.Based on the application requirements,an urban greening plant monitoring scenario consisting of 17 plants and a field and economic crop monitoring scenario consisting of 16 plants were determined for subsequent analysis.For the monitoring of plant growing status,several control experiments were carried out for tea plant growth monitoring,wheat stripe rust severity monitoring and rice growth stage monitoring.In addition,an UAV hyperspectral image was obtained for a test of corresponding methods.These experiments provide a basis for subsequent plant spectral feature extraction and studies on classification algorithm.(2)Study the spectral feature extraction and preferred methods.Various types spectral features were extracted including spectral differentiation,continuum and vegetation index.A feature sensitivity analysis is proposed by combining ISODATA band clustering and JM distance,and the preferred spectral features are obtained for different monitoring scenarios.Besides,a spectral feature robustness analysis method is proposed by simulating noise and illumination effects.It is found that the spectral bands around 400nm,700nm,1500nm,2000nm and PVIhyp,NDWI,and sLAIDI are sensitive in plant classification modeling.Spectral band and red edge differential features around 700nm,1000nm,1400nm ? 2000nm,BRI,LWVI-1 and other spectral features are sensitive to plant growing status monitoring.(3)Study hyperspectral plant classification and growing status monitoring algorithms.The algorithms included K-based neighbor(KNN)algorithm,genetic algorithm couplingwith support vector machine(GA-SVM),and the non-parametric random forest(RF)algorithm The results show that the GA-SVM algorithm performs best in plant classification.The highest accuracy in urban greening plant monitoring scenarioreaches OAA=0.98 and Kappa=0.98.The highest classification accuracy in the field and economic crop monitoring scenario is OAA=0.99 and Kappa=0.99.In the monitoring of plant growing status,the GA-SVM algorithm also performed the best in tea plant growth monitoring(OAA=0.91,Kappa=0.86)and monitoring of rice growing stages(OAA=0.93,Kappa=0.92);but RF performed the best(OAA=0.65 and Kappa=0.46)monitoring in wheat stripe rust.Among various scenarios and applications,GA-SVM performed the best,following by RF algorithm,and then the KNN algorithm.The feature selection and classification algorithm was further performed on the UAV hyperspectral image data,which yielded an accuracy as OAA=0.87,Kappa=0.83.(4)Development of a spectral library system for vegetation.Based on the experiments and methods of plant classification and monitoring of plants' growing status,this study established a spectral library system for vegetation,which achieved the storage and retrieval of spectral features,sensitivity analysis,and plant classification and growing status monitoring.Key words:hyperspectral;plant classification;plant growing status monitoring;feature extraction;sensitivity analysis;classification algorithm. |
PDF Full Text Request