| Obtaining the biophysical and biochemical information of alpine grassland efficiently and accurately lays a foundation to sustainable utilization of grassland resources and scientific development of grassland animal husbandry.The traditional chemical analysis method is time-consuming and costly,multi-spectral remote sensing is limited by spectral channels and discontinuous wide bands,which is generally constrained by the macro research of grassland biophysical indicators.The rapid development of hyperspectral remote sensing technology and its widespread application in monitoring of vegetation biochemical parameters provide new opportunities and challenges for estimating the key nutrients and important growth parameters of alpine grassland.Taking alpine grassland in the east of Tibetan Plateau as the research object,based on the grassland field observation experiments for five years(2014-2018),using hyperspectral remote sensing,mathematical statistical analysis and machine learning modeling,this study analyzes the forage hyperspectral remote sensing characteristics at different growth stages of alpine grassland and achieves the remote sensing estimation of the forage key nutrients(i.e.,nitrogen(N)and phosphorus(P))and the important evaluation indexes of the growth status(i.e.,carbon-nitrogen ratio(C:N)and nitrogen-phosphorus ratio(N:P)).The purpose of this study is to improve the hyperspectral remote sensing monitoring method of forage biochemical parameters,subsequently,providing a theoretical basis for the rational utilization of grassland resources and the balance of forage and livestock nutrition.The main results show that:(1)At the vegetative growth stages(from the early period of forage growth to the vigorous growth period),the spectral reflectance of the forage canopy in the visible region decreases gradually,and the reflectance in the near-infrared(NIR)region increases significantly due to multiple scattering.As the forage gradually senesces(from the vigorous growth period to the senescence stage),the absorption of the visible region by the forage gradually weakens,the reflectance in the visible region increases while the NIR region decreases significantly.When the forage is completely senesced in November,the spectrum is similar to the spectral characteristics of the soil,showing a slow increasing trend within the wavelengths from 350-1350 nm without obvious absorption or reflection features.On the whole,throughout the growth period,the red-edge first shifts toward longer wavelengths and then shifts toward shorter wavelengths,the amplitude(AMP)and absorption depth(AD)gradually decrease,and the absorption position(AP)changes slightly.(2)The applicability of spectral variables for forage N estimation differs during different growth periods,and the multivariate model achieves better estimation accuracy during the growth and reproduction stages of forage than the withering period(R~2:0.58-0.68 versus0.23).Among the 38 spectral variables composed of vegetation indexes,absorption bands,red-edge parameters and absorption features,some spectral variables(i.e.,NDNI,VOG,PRI,SIPI,and REP)that are sensitive to chlorophyll and N play a critical role in estimating forage N during the vegetative growth and reproductive growth periods of forage.However,some spectral variables,such as SAVI,OSAVI,BDR,and NBDI,may contribute significantly to detecting N during the withering period of forage.(3)The combination of hyperspectral bands and multiple-factors(e.g.,geography,topography,soil,vegetation and meteorology)can improve the estimation accuracy of forage P in the alpine grassland ecosystem of the Tibetan Plateau.First derivative(FD)and continuum removal(CR)spectra can retrieve more feature bands that are mainly located in the NIR and shortwave infrared(SWIR)regions than original spectra for the forage P.Some factors,such as Longitude,Elevation,Sand1,Clay1,Sand1/Clay1,and Temperature,have extremely significant correlations with forage P(R?0.196,P<0.01),and the Longitude and Temperature are important for forage P estimation.The optimum forage P inversion model(R~2=0.67,RMSE=0.0472%)established by the support vector machine(SVM)algorithm and FD spectral bands can account for 88%of the variation of forage P in alpine grassland.(4)Some known feature absorption bands(KBs)associated with protein,chlorophyll,N and carbon-containing compounds exhibit good performance in estimating the forage C:N ratio(R~2 of 0.70-0.80).The important bands(IBs)derived from the red-edge and red regions significantly contribute to the forage C:N ratio estimation with R~2 values of 0.77 to 0.80.The combination of KBs and IBs(CBs)can slightly improve the prediction accuracy of the forage C:N ratio compared to using IBs and KBs alone(an increase in R~2 values of 0.01-0.02),and further optimization of CBs can avoid the saturation effect of the model.The optimum forage C:N inversion model presents satisfactory estimation accuracy(R~2=0.82,RMSE=5.5308),explaining 85-92%of the variation in the forage C:N ratio during different growth stages(May to November).(5)It is practical and feasible to use Sentinel-2 MSI spectral bands and vegetation indices to estimate the forage N:P ratio.During the vigorous growth period of forage,NDII and RECI2 significantly contribute to the forage C:N ratio estimation,and the bands sensitive to forage N:P ratio estimation are mainly distributed in NIR and SWIR region.During the withering period of forage,in addition to the bands of NIR and SWIR region,some bands located in red and red-edge region also have a significant influence on forage N:P ratio estimation,and NDWI,RECI2,NDRE1 and NDRE2 are the most important vegetation indices for estimating forage N:P ratio.Compared with the forage N:P ratio model established by using the optimized spectral bands and vegetation indices alone,the estimation accuracy(R~2)of the model constructed by combining the above two types of variables is 0.49 and 0.59,which is improved by 0.06 and 0.04 in the vigorous growth period(July)and the withering period of forage(November),respectively. |
PDF Full Text Request