Estimation Grassland Above Ground Biomass Based On UAV Technology And Machine Learning Methods In Alpine Grassland,Gannan Region

Above ground biomass?AGB?is an important indicator of regional carbon cycling,grassland ecosystem health assessment and grassland resources sustainable utilization.Grassland biomass of the Tibetan Plateau is considered as an indicator of global climate change,due to the special climatic environment and complex topographic conditions,which mostly reflect the dynamic changes in the vegetation species evolution and ecosystems development in the plateau.Therefore,construct an accuracy grassland biomass estimation model is particularly important for grassland management,livestock balance,grassland growth status assessment and ecological environment protection.Several grassland biomass estimation researches have been done in the Tibetan Plateau.However,the biomass estimation models differed greatly,and most suitable biomass estimation model for the grassland biomass is currently unclear.This study was conducted as a case study in the Gannan region to investigate the following content:constructthesingle-factorparametricandmulti-factor parametric/non-parametric biomass estimation models based on the factors?geographical location,topography,grassland biophysical,meteorological,soil,remote sensing vegetation index?,which have significant correlation with grassland biomass.Then compare the accuracy and stability of each model by 10-fold cross validation method.Finally,based on the results above,analyze the temporal and spatial dynamic of grassland biomass from 2000 to 2016 in the study area.The results showed that:1)The filtering and de-noising processing of MOD13Q1 indices are the key methods for reducing the AGB inversion error of alpine meadow grassland based on MODIS data.The accuracies of the optimal estimation models based on the EVI,which filtered by Gaussian,Double-logistic and Savitzky-Golay,are higher than MOD13Q1 EVI,the R² increased 0.001⁰.009 and the RMSE decreased 0.72 kg/ha¹.33 kg/ha in test data set.Besides,the accuracies of optimal estimation models based on NDVI,which filtered by Gaussian,Double-logistic and Savitzky-Golay,are higher than MOD13Q1 NDVI,the R² is increased 0.057⁰.068,and the RMSE is reduced 19 kg/ha³4.21 kg/ha of the test data set.2)Nine of the 14 examined factors presented significant correlations with grassland biomass.X,Y,DEM,TPI,H,C,T,P,and Clay2 were significantly or extremely significantly correlated with grassland biomass.In all optimal biomass estimation models based on each factors,the model based on C has the highest precision with R² of 0.39 and RMSE of 859.368 kg/ha,respectively.The following are based on H,P,T,Y,DEM,X,TPI,and Clay2.Overall,grassland estimation models based on a single factor exhibited poor accuracy and low stability.The single-factor grassland biomass estimation models accounted for only 1.5%³6.3%of the variation in biomass.3)An important method of improving the accuracy of biomass estimation is to construct multi-factor parametric/non-parametric estimation models.The trend of accuracy changes are similar in the multi-parameter and non-parametric models based on each factors.Among all factors,the precision of optimal model based on vegetation physical indices?C and H?is the highest in two types of models;followed by the model based on meteorology?T and P?,geographic location?X and Y?,and topography?DEM and TPI?.However,the accuracy of the non-parametric models constructed by the above four factors is higher than that of parametric models,with the R² increased 0.07⁰.1,and the RMSE decreased 1.41 kg/ha⁸0.65 kg/ha.The trend of the accuracy changes are similar in parametric and non-parametric models based on each combination factors,with the input factors reduction,the accuracy of two types of models is gradually decreasing.The accuracy of the non-parametric models is higher than parametric models in the same combinations,with R² increased0.04⁰.12,and RMSE reduced 74.91 kg/ha¹38.34 kg/ha.4)An effective method in improving the stabilities of biomass estimation models is through the construction of multi-factor parametric/non-parametric estimation models.In this study,the stability is better than single-parametric models based on each factor,either verified by the multi-parametric or non-parametric models based on each combination factors.With the factors increased,the stability of the models increased both in multi-factor parametric and non-parametric models.The stability of optimal non-parametric model based on combination of X,Y,DEM,TPI,H,C,T,P,Clay2 and EVI_SG is the highest,compared to the sing-/multi-factor parameter models,the SD_RMSEMSE reduced 37.508 kg/ha⁶1.286 Kg/ha.5)Grassland biomass showed an increased trend in Gannan region from 2000 to2016,with an increase of 12.463 kg/ha per year.For the spatial distribution of grassland biomass changes,the trend of changes in biomass mainly in a stable and increased state,with 78.53%and 20.50%of the entire grassland area,respectively.Only 0.98%of the study area is in decreased state.For each grassland types in study area,the mainly trend of changes also in a stable and increased state,except for low-lying meadow grassland,with 76.81%⁹7.64%in stable and 12.55%²3.02%in increased state,while the proportion of reduced area is only 0.17%⁶.43%.