Remote Sensing Estimation Of Grassland Fractional Vegetation Cover (FVC) Based On Machine Learning Method In Qilian Mountain

Grassland fractional vegetation cover(FVC)is a key indicator to measure the growth status of surface vegetation,it is also an indicator for regional carbon cycle analysis,grassland ecosystem health assessment and sustainable development and utilization of grassland and livestock resources.Qilian Mountain is a typical ecologically fragile region in northwestern China,and grassland ecosystem dominates in this region.Therefore,building a high-precision estimation model of grassland FVC is of great significance to the livestock balance analysis,ecological environmental protection in Qilian Mountain.A number of researchers have focused on the estimation of FVC in Qilian Mountain.However,current models have great difference in estimating different types of vegetation FVC,there still lack a systematical and comprehensive research in the estimation of grassland FVC in the eastern,middle and western regions of Qilian Mountain.Therefore,this thesis took Qilian Mountain as a typical research area,and the various types of grassland vegetation in the eastern,middle and western regions were regarded as research objects,then compared the applicability of different MODIS products to estimate the grassland FVC in this region.Various vegetation indices,bare soil index,shadow index and meteorological factors such as temperature,precipitation and humidity were used as input data.Single-factor and multi-factor parametric and non-parametric models of grassland FVC were conducted based on these input data,then compared the accuracy together with stability of all models.Based on them,the spatial-temporal dynamics of grassland FVC in Qilian Mountain from 2000 to 2019 were analyzed,and the results show that:(1)Compared with MOD09 GA,MCD43A4 is more suitable for the research of grassland FVC in Qilian Mountain.Among 12 types of vegetation indices,the determination coefficients between grassland FVC and 11 kinds of VIs based on MCD43 are higher than those VIs based on MOD09,and the difference in r of the same VI was between 0.031~0.793.(2)Due to the impacts of topography,soil,climate and different types of vegetation cover,the ability of single-factor models based on different indices to estimate the grassland FVC in the eastern,middle and western sections in Qilian Mountain have great differences.The optimal model of grassland FVC in the eastern part of Qilian Mountain is the exponential model based on BNDVI(y = 19.7160)1.292,R2=0.569,RMSEP=1.582),the optimal model of grassland FVC in the middle section of Qilian Mountain is the power model of GNDVI(y = 65.5300.543,R2=0.698,RMSEP=15.601),the optimal estimation model of grassland FVC in the western section of Qilian Mountain is the power model of DFI(y = 145.8681.068,R2=0.451,RMSEP=4.573).(3)To construct multi-factor parameter models is an effective method for the improvement of the accuracy of grassland FVC inversion.The multi-factor model using composite vegetation indices,soil index,shadow index together with meteorological factors as input factors has the best accuracy of the estimation of grassland FVC in the eastern,middle and western sections of Qilian Mountain.The R2 of their multivariate linear models is 1.000,0.730 and 0.885 respectively,and the RMSEP is 3.485,11.573 and 4.375 respectively.The accuracy of the multifactor parameter models in the middle and western sections in Qilian Mountain has been improved by 27.25% and 68.49% respectively with the addition of meteorological factors in the models,which is higher than the improvement of the accuracy by the soil index and shadow index to the models(3.92% and 26.42% respectively).(4)The non-parameter models based on BPNN method have great improvement of the accuracy of single-factor and multi-factors which to estimate the grassland FVC in Qilian Mountain.In the middle section,the R2 of the single-factor optimal estimation model based on BPNN of 3 types of VIs(BNDVI,GNDVI,NDVI)has increased by 0.168,0.131 and 0.055 respectively,and the RMSEP has decreased by 11.602,8.343 and 8.939.In the western section,the R2 of the single-factor optimal estimation model based on BP-ANN of 3 VIs types(DFI,NDVI,MSAVI)has increased by 0.322,0.422 and 0.300 respectively,and the RMSEP has decreased by 3.312,2.855 and 6.562.Based on the same input data,compared with the multifactor linear model,the R2 and RMSEP of BPNN model have improvement in different degrees in the middle and western section in Qilian Mountain.(5)The grassland FVC from west to east of Qilian Mountain exhibits a gradual increasing trend.The average grassland FVC in the eastern,middle and western sections in Qilian Mountain are 49.734%,41.035% and 18.015% respectively.Overall,the grassland FVC in Qilian Mountain has shown an upward trend since 2000 to 2019.The average increase in eastern,middle and western sections in Qilian Mountain is about 4.222%,4.381% and 2.673%.From the spatial change trend of grassland FVC,the grassland FVC in Qilian Mountain exhibits a non-significant increase trend from 2000 to 2019,and the grassland area which did not increase significantly accounted for 80.32% of total area.From the perspective of different grassland types,the grassland FVC of each type is mainly because of the state of non-significant increasing,and the proportion of their area in each type of grassland ranged from 68.775% to 94.141%.