Study On Optimal Feature Extraction And Model Development From The Hyperspectral Image For Leaf Biomass In Wheat

Quantitative monitoring of crop biomass is an important research topic in the field of vegetation remote sensing.It is critical to estimate crop biomass for assessing crop growth and guiding fertilization in precision agriculture.The hyperspectral techniques provide fast,non-destructive and powerful approaches to monitor crop biomass.This study obtained hyperspectral images of wheat canopy by imaging spectrometers(ImSpector V10E-PS,Spectra Imaging Ltd,Finland),as well as extracting the wheat leaf biomass by destructive sampling on the basis of multiple field experiments under two years,varied N rates,planting densities and cultivars in wheat.The primary objective of this study is to explore the preprocessing methods for hyperspectral images.In order to extract the optimal hyperspectral features of wheat leaf biomass from full wavelengths,combing synergy interval partial least squares with successive projections algorithm(SIPLS-SPA)is proposed.Then extreme learning machine(ELM)is employed to establish quantitative monitoring model of wheat leaf biomass.The anticipated results will provide new wavebands choices for not only manufacturing portable crop growth instrument but also utilizing space-borne remote sensing data.Therefore,the real-time estimation and precise diagnosis of crop biomass in wheat can be realized.Firstly,the wheat canopy hyperspectral reflectance under different N rates and the linear correlation of wheat leaf biomass to hyperspectra reflectance are identified.Then the hyperspectra within the range of 400-1000 nm are systematically analyzed to determine the characteristic hyperspectra by employing SIPLS-SPA.In addition,compared with the hyperspectral features selected by SIPLS-SPA,synergy interval partial least squares(SIPLS)and synergy interval partial least squares(SPA)are adopted separately to extract hyperspectral features.SIPLS-SPA selects eight hyperspectral variables in 706,724,734,806,808,810,812 and 816 nm,which are used as input variables of PLSR and provids R2c of 0.80,R2v of-0.06,RMSEv of 0.059(kg/m2)and RRMSEv of 38.66%,respectively.The hyperspectral features extracted by SIPLS-SPA can represent the full spectral information for estimating wheat leaf biomass.Comparing with SIPLS,SIPLS-SPA provids fewer unrelated and collinear spectral variables and significantly simplify the calibration model.Comparing with SPA,SIPLS-SPA also improves the accuracy and stability of the calibration model.In this study,we established estimation model of wheat leaf biomass in the whole growing season with the optimal hyperspectral features by employing extreme learning machine(ELM).To fully highlight the advantages of ELM method,traditional methods(PROSAIL method,vegetation indices method,partial least squares(PLS),neural network(NN)and support vector machine(SVM))are also employed to compare with ELM method.The results demonstrate that ELM model offers the best accuracy in terms of R2c,R2v,RMSEv and RRMSEv at 0.78,0.57,0.038(kg/m2)and 24.57%,respectively.Comparing with vegetation index and PROSAIL methods,ELM method improves the accuracy of estimation model.Comparing with NN,ELM is simpler in topology structure and ran faster.Comparing with SVM,ELM has a simpler principle and less parameters.Compared with PLS and vegetation indices,ELM is more suitable for analyzing the non-linear relationships between heavy biomass and hyperspectral variables.Overall,ELM method improves the reliability,applicability and practicability of eatimation model.