| Desert steppe of Ordos is an important part of grassland in Inner Mongolia.In recent years,due to natural and man-made factors in the region,problems such as degradation of grassland area and decrease of production capacity have resulted in serious impacts on the sustainable development of the local ecological environment.Taking effective monitoring of grassland degradation and specifying pasture grassland yield are important guarantees for forage and animal balance,prevention of grassland degradation and protection of grassland resources.The hyperspectral remote sensing technique is used to estimate the biophysical parameters of the grassland,which provides a scientific basis for the remote sensing monitoring of the Ordos desert steppe and the scientific management of the grassland.Because of different ecological regions and different degradation gradients,the canopy structure and spectral reflectance of grassland vegetation are different,which limits the accuracy of hyperspectral remote sensing estimation model.Therefore,this paper took the desert steppe of Otog Banner in Ordos,Inner Mongolia as the research object,and used hyperspectral remote sensing technology to carry out field investigation,biophysical parameter and ground hyperspectral spectroscopy determination in Ordos desert steppe in July and August from 2014 to 2016,analyzed spectral characteristics of grassland canopy,typical plant community(Caragana stenophylla community,Stipa breviflora community,Cleistogenes songorica community,Artemisia frigida community)canopy and typical plant(Caragana stenophylla,Stipa breviflora,Cleistogenes songorica,Artemisia frigida)canopy in Control,Light Degraded,Medium Degraded,Heavy Degraded and All-areas,analyzed the response of hyperspectral characteristics and degradation gradient of desert steppe,explored the relevance of primitive spectrum,differential spectrum,vegetation indices,hyperspectral characteristic parameters and biophysical parameters of grassland canopy,typical plant community canopy and typical plant canopy,extracted the sensitive characteristic bands,vegetation indices and hyperspectral characteristic parameter,and used linear and non-linear of single-variable,stepwise regression and BP neural network methods to establish hyperspectral estimation model of biophysical parameters of desert steppe,providing a theoretical basis and technical support for the Ordos desert steppe growth conditions and dynamic monitoring.The main findings were as follows:(1)The primitive,first-derivative and second-derivative spectrum of grassland canopy,typical plant community canopy and typical plant canopy in the Ordos desert steppe along the same and different degradation gradients were analyzed.(1)The primitive spectral reflectance of grassland canopy,typical plant community canopy and the typical plant canopy in Ordos desert steppe was low in the visible wavelength band of 380–700nm,not more than 15%,and high in the visible wavelength band of 780–1830nm,but not more than 35%.(2)In the visible wavelength band,the primitive spectral reflectance of four degraded desert steppe was:HD>MD>LD>CK;in the near infrared band,the relationship of the four was:CK>LD>MD>HD.The first-derivative and second-derivative spectrum of grassland canopy were consistent along different degradation gradients,and the first-derivative maximum value was around 719nm at the red edge position.(3)The primitive,first-derivative and second-derivative spectrum of the typical plant community were consistent along different degradation gradients.The primitive spectral reflectance of community canopy of Caragana stenophylla increased in the visible wavelength band of 350–680nm with the degree of grassland degradation.The primitive spectral reflectance of community canopy of Stipa breviflora and Artemisia frigida decreased in the visible wavelength band of 780–1350nm with the degree of grassland degradation.There was no obvious change in the primitive spectrum of community canopy of Cleistogenes songorica.The first-derivative spectrum maximum values of all the typical plant communities are at 719 nm,with the order of CK>LD>MD>HD.The second-derivative spectrum has no obvious changes near 719nm.(4)The primitive,first-derivative and second-derivative spectrum of the typical plant are consistent along different degradation gradients.The primitive spectral reflectance of plant canopy of Caragana stenophylla and Artemisia frigida increased in the visible wavelength band of 350–680nm with the degree of grassland degradation.The primitive spectral reflectance of plant canopy of Caragana stenophylla,Cleistogenes songorica and Artemisia frigida decreased in the visible wavelength band of780–1350nm with the degree of grassland degradation.The first-derivative spectrum maximum values of typical plants were around 719nm.The order of first-derivative reflectance of Caragana stenophylla,Cleistogenes songorica and Artemisia frigida were CK>LD>MD>HD.There was no obvious change near 719nm in second-derivative spectrum.(5)The spectral reflectance of the typical plant communities in CK and LD was Cleistogenes songorica community>Artemisia frigida community>Stipa breviflora community>Caragana stenophylla communities in the primitive spectrum wavelength band of 350–680nm along the same degradation gradient.With the increase of grassland degradation,the reflectivity of Cleistogenes songorica community was the largest.There were no obvious changes in Caragana stenophylla,Stipa breviflora,and Artemisia frigida community.There was no obvious change of the typical plant communities in the primitive spectrum wavelength band of 780–1350nm.The first-derivative and second-derivative spectrum of the typical plant community canopy showed a significant difference near 719nm along the same degradation gradient.(6)The typical plant spectral reflectance in CK was Stipa breviflora>Cleistogenes songorica>Artemisia frigida>Caragana stenophylla in the primitive spectrum wavelength band of 350–680nm,and Cleistogenes songorica>Stipa breviflora>Artemisia frigida>Caragana stenophylla in LD,MD,and HD along the same degradation gradient.There was no obvious change in the primitive spectrum wavelength band of 780–1350nm.The first-derivative and second-derivative spectrum of the typical plant canopy showed a significant difference near 719nm along the same degradation gradient.(2)The optimized normalized vegetation indices(NDSI,FNDSI and SNDSI),ratio vegetation indices(RSI,FRSI and SDSI)and differential vegetation indices(DSI,FDSI and SRSI)were constructed by the primitive spectrum,first-derivative spectrum and second-derivative spectrum of all the possible two-band combinations in full wave-band.Based on the correlation analysis between the optimized vegetation indices and the biophysical parameters of desert steppe,the sensitive vegetation indices were found,which laid the foundation for the establishment of the best estimation model of biophysical parameters.(3)The correlation between the primitive spectrum,the differential spectrum,the vegetation indices,the hyperspectral characteristic parameter and the grassland fresh biomass along different degradation gradients and in ALL-areas were analyzed to determine the sensitive spectral variables.This paper used linear and non-linear of single-variable,stepwise regression and BP neural network to establish and compare grassland fresh biomass estimation model along different degradation gradient and in ALL-areas,and determined the optimal estimation model.(1)The correlation of first-derivative spectrum and grassland fresh biomass of grass canopy along different degradation gradient and in ALL-areas was best,and the best bands were 559,712,710,703 and 710nm.(2)The optimal vegetation indices along different degradation gradient and in ALL-areas were NDSI(706,707)and RSI(707,706),NDSI(582,719),NDSI(708,710)and RSI(710,708),FRSI(499,703),NDSI(705,710)and RSI(710,705).(3)The hyperspectral characteristic parameters with significant correlation of grassland fresh biomass along different degradation gradient and in ALL-areas were Dy,DR,Rg,Rr,SDy,SDR,Rg/Rr,(Rg-Rr)/(Rg+Rr),SDR/SDb,(SDR-SDb)/(SDR+SDb)and(SDR-SDy)/(SDR+SDy).(4)Comparing grassland fresh biomass estimation model established by different spectral variables and modeling methods along different degradation gradient and in ALL-areas,the optimal estimation model was BP neural network estimation model which established based on the vegetation indices.The fitted power function model R~2 were0.736,0.785,0.808,0.691,0.773,the verification power function model R~2 were 0.815,0.828,0.866,0.613,0.787,RMSE were 40.55,36.31,21.48,33.76,38.47 g/m~2,and RE were 28.70,20.04,18.85,30.98,29.48%.It can be seen from the analysis that using the corresponding estimation model can estimate the fresh biomass of grassland more accurately along different degradation gradients.MD has the highest estimation accuracy and the HD estimation accuracy is relatively low.(4)The relationship between primitive spectrum,differential spectrum,vegetation indices,hyperspectral characteristic parameters and fresh biomass of typical plant community canopy was analyzed.The sensitive spectral variables were determined.Linear and non-linear of single-variable,stepwise regression and BP neural network were used to establish estimation model along different degradation gradient and in ALL-areas and the optimal estimation model was established.(1)The best characteristic bands related to the fresh biomass of typical plant communities were mainly located at the red edge position,followed by the yellow edge position along different degradation gradient and in ALL-areas.(2)The best combination of vegetation indices with the fresh biomass of typical plant communities was mainly located in the visible bands of primitive and differential spectrum along different degradation gradient and in ALL-areas.(3)The parameters with large correlation coefficients of typical plant community fresh biomass along different degradation gradient and in ALL-areas were Dy,Rg,Rr,SDy,SDR,Rg/Rr,(Rg-Rr)/(Rg+Rr),SDR/SDb,(SDR-SDb)/(SDR+SDb)and(SDR-SDy)/(SDR+SDy).(4)By comparing the different spectral variables and different modeling methods to establish typical plant community fresh biomass estimation model along different degradation gradient and in ALL-areas,the optimal estimation model was BP neural network estimation model which established based on the vegetation indices.For Caragana stenophylla community fresh biomass,the fitted power function model R~2were 0.856,0.863,0.936,0.844,0.793,the verification power function model R~2 were0.896,0.921,0.968,0.822,0.833,RMSE were 36.93,30.81,16.61,30.10,49.05g/m~2,and RE were 12.57,10.45,14.18,19.66,23.24%.For Stipa breviflora community fresh biomass,the fitted power function model R~2 were 0.818,0.854,0.937,0.925,0.675,the verification power function model R~2 were 0.847,0.883,0.830,0.828,0.732,RMSE were 33.90,20.05,21.48,20.66,40.80g/m~2,and RE were 17.58,16.50,23.64,21.16,28.07%.For Cleistogenes songorica community fresh biomass,the fitted power function model R~2 were 0.821,0.964,0.949,0.801,0.818,the verification power function model R~2 were 0.913,0.989,0.948,0.812,0.841,RMSE were 19.23,9.65,9.82,10.25,27.58g/m~2,and RE were 15.37,11.96,13.98,21.83,22.50%.For Artemisia frigida community fresh biomass,the fitted power function model R~2 were 0.963,0.972,0.943,0.824,0.810,the verification power function model R~2 were 0.980,0.974,0.974,0.885,0.822,RMSE were 15.68,15.45,10.90,20.02,40.21g/m~2,and RE were 10.00,9.52,13.34,19.03,25.72%.It can be seen from the analysis that typical plant community fresh biomass estimation model based on different community grassland types can improve the estimated accuracy of Ordos desert steppe.Estimation model,estimation accuracy and verification effects of typical plant community fresh biomass established along different degradation gradient were superior to that of in ALL-areas.(5)Along different degradation gradient and in ALL-areas,the relationship between typical plant canopy hyperspectral data and plant water content was analyzed.The sensitive spectral variables were determined.Linear and non-linear of single-variable,stepwise regression and BP neural network were used to establish and compare estimation model along different degradation gradient and in ALL-areas and the optimal estimation model was established.(1)The best characteristic bands of the typical plant canopy spectrum and plant water content were located at the yellow edge position,followed by the red edge position.(2)The best vegetation indices combination of plant canopy spectrum and plant water content of Caragana stenophylla was mainly located in the near infrared band of primitive and differential spectrum,while Stipa breviflora,Cleistogenes songorica and Artemisia frigida were located in the visible bands.(3)The correlation changes greatly between the hyperspectral characteristic parameters and plant water content of typical plant canopy along different degradation gradient and in ALL-areas.(4)By comparing the different spectral variables and different modeling methods to establish estimation model of typical plant water along different degradation gradient and in ALL-areas,the optimal estimation model was BP neural network estimation model which established based on the vegetation indices.For Caragana stenophylla plant water content,the fitted power function model R~2 were 0.753,0.781,0.867,0.740,0.450,the verification power function model R~2 were 0.876,0.612,0.850,0.676,0.437,RMSE were 4.57,4.36,3.01,4.51,7.61%,and RE were 10.46,6.68,4.38,8.93,29.36%.For Stipa breviflora plant water content,the fitted power function model R~2 were 0.850,0.882,0.883,0.796,0.614,the verification power function model R~2 were 0.843,0.846,0.878,0.815,0.736,RMSE were 3.09,2.77,2.51,2.80,4.04%,and RE were 6.90,5.82,6.50,6.36,9.95%.For Cleistogenes songorica plant water content,the fitted power function model R~2 were 0.789,0.877,0.949,0.946,0.542,the verification power function model R~2 were 0.827,0.816,0.966,0.976,0.690,RMSE were 3.59,3.68,2.84,1.37,6.00%,and RE were 4.32,5.41,3.39,2.06,8.54%.For Artemisia frigida plant water content,the fitted power function model R~2 were 0.810,0.866,0.752,0.900,0.766,the verification power function model R~2 were 0.927,0.842,0.960,0.935,0.802,RMSE were 2.78,3.48,2.03,3.60,5.05%,and RE were 3.97,4.49,2.93,8.80,8.76%.In summary,it provides a theoretical basis and technical support for the desert steppe degradation control and dynamic monitoring by using hyperspectral technology to establish the optimal estimation model for the plant fresh biomass,typical community fresh biomass and typical plant water content of grassland of Ordos desert steppe along different degradation gradient and in ALL-areas based on vegetation indices BP neural network and effectively improve the estimation accuracy. |
PDF Full Text Request