| Grassland resources are an important part of the global terrestrial ecological cycle system and one of the most common and widely distributed terrestrial vegetation types,accounting for 41.7%of the total land area of the world.Grasslands are also the largest carbon source on land except for forests,and they play a vital role in regulating the global carbon cycle and supporting biodiversity of plants and animals.In agricultural production,natural grassland not only provides the forage for animal husbandry in the most primitive way,but provides energy and fiber for ruminants as well.The nutrient of herbage also determines the cost and quality of the production in animal husbandry,which is an important index to evaluate the value of grassland resources.Therefore,the rapid analysis and evaluation of the nutrient of grassland and herbage can effectively provide primary data for grassland management and forage budget.The evaluation also has guiding significance for the protection of ecological resources and the development of animal husbandry.From the perspective of precision agriculture,this study took Leymus chinensis,the most common grassland resources in northeast area of china,as research objects.The whole growing period data was collected by month and taken advantages of to establish a combined database about forage grass nutrition content and spectral information from lots of remote sensing platforms.A variety of spectral feature extraction methods were used to make comparison with the model of quantitative analysis.The inversion study was conducted on four kinds of nutrients of forage grass:dry matter,crude protein,acid washing fiber and neutral detergent fiber from three dimensions of the interior vane,unmanned aerial vehicle(UAV)canopy and satellite area.The research randomly placed 35 quadrats as samples in the artificial pasture fields on May 17,June 20,July 18,August 18 and September 17,2017 respectively.With the help of the UAV with multi-spectral camera,the above 30 meters of the sample point were shot in the normal direction,and then the above ground parts of all the herbage in the sample square were collected.Transferred to the laboratory,the samples were used to collect the hyperspectral data of the leaves and to determine the chemical values of the nutrients.High-quality sentinel 2 satellite remote sensing data was selected for a close date to the sampling date each month.The nutrient composition of forage grass was modeled on three scales and inversed on regional scales.On indoor blade scale,comparison was made to identify the influences made by five kinds of pretreatment methods(the Savitzky-Golay convolution smoothing(SG),multiple scatter correction(MSC),variable standardization(SNV),the first derivative(1-Der)and direct orthogonal signal correction(DOSC))to the partial least-squares regression(PLSR)prediction model of nutrient content.The results showed that the optimal pretreatment method of forage dry matter(DM),crude protein(CP),and both neutral detergent fibers(NDF)and acid detergent fibers(ADF)is 1-Der,SNV and DOSC.By adopting continuous projection algorithm(SPA),genetic algorithm(GA),competitive weighted adaptive algorithm(CARS)and random frog algorithm(RF),the characteristic bands of four nutrients in forage grass were selected.And the results showed that the sensitive bands of dry matter in forage grass were mostly distributed in the ranges of 450-460nm,750-780nm and 840-900nms and 530nm-700nm and 940nm-1000nm for the crude protein,450-470nm,570-590nm and 610-630nm for neutral detergent fibers and 580-620nm,920-950nm for the acid detergent fibers.After establishing the PLSR models based on the characteristic wavelength screened by different algorithms,the optimal characteristic variable selection method for the dry matter of forage grass was SPA,and the R2-P,RMSEP and RPD of PLSR prediction model were 0.927,35.852 and 4.257 respectively whereby the optimal prediction model was 1-der-SPA-PLSR;For crude protein,RF was the best and the R2-P,RMSEP and RPD of PLSR prediction model were 0.933,6.034 and 4.322 respectively whereby the optimal prediction model was SNV-RF-PLSR;For neutral detergent fibers(NDF),GA was the best and the R2-P,RMSEP and RPD of PLSR prediction model were 0.797,16.357 and 2.425 respectively whereby the optimal prediction model was DOSC-GA-PLSR;Accordingly,for acid detergent fibers(ADF),SPA was the best and the R2-P,RMSEP and RPD of PLSR prediction model were 0.768,14.985 and 2.280 respectively and the optimal prediction model was DOSC-SPA-PLSR.In the study of UAV canopy scale,the multi-spectral data obtained by UAV were firstly analyzed to construct 19 vegetation indexes.Then the correlation analysis(CA)and stepwise regression analysis(SR)of forage nutrient composition were carried out by using 28 spectral characteristic variables.And finally,a selection of the 28 spectral characteristic variables was made.The correlation coefficient method was used to select 7 spectral characteristic variables with the highest correlation with the content of each nutrient component.Through stepwise regression method,the number of the spectral characteristics in relation to the model determined for the DM,CP,NDF and ADF was 5,5,3 and 3 respectively.These spectral characteristics were used in modeling forage nutrient content by using multiple linear regression(MLR),partial least squares regression(PLSR),least squares support vector machines(LS-SVM)and the extreme learning machine(ELM)separately and a comparison of the models' performances was made.Results showed that:the optimal prediction model of forage dry matter(DM)on UAV canopy scale was CA-PLSR and the R2-P,RMSEP and RPD of the prediction model were 0.734.16.164 and 2.993 respectively;Accordingly,the optimal prediction model of crude protein(CP)was SR-PLSR and the R2-P,RMSEP and RPD of the prediction model were 0.734.16.164 and 2.993 respectively;The optimal prediction model of neutral detergent fibers(NDF)was CA-ELM and the R2-P,RMSEP and RPD of the prediction model were 0.592,24.351 and 1.554 respectively;The optimal prediction model of acid detergent fibers(ADF)was CA-LS-SVM and the R2-P,RMSEP and RPD of the prediction model were 0.608,20.045 and 1.552 respectively.In the satellite regional scale,after extracting the grassland distribution of objective area based on unsupervised classification,the spectrum curve of 8 bands of Sentinel-2 was analyzed.The three vegetation indices which had high correlations was selected based on the correlation analysis results of vegetation index and nutrient composition at canopy scale.The vegetation index was constructed and selected based on Sentinel-2 data by combining adjacent bands.The four vegetation indices with high correlation to the nutritional ingredients were OSAVI(B8,B4),NDRE(B8,B6),MCARI(B8,B4,B3)and CIred-edge(B8,B5).The empirical regression model was builded by the four vegetation indices,included linear model,exponential model,logarithmic model,quadratic model.The optimal prediction model of DM was quadratic model by OSAVI(B8,B4);The optimal prediction model of CP was logarithmic model by NDRE(B8,B6);The optimal prediction model of NDF was logarithmic model by MCARI(B8,B4,B3);The optimal prediction model of ADF was exponential model by CIred-edge(B8,B5);... |
PDF Full Text Request