| Vegetation is an indispensable and important role of the ecosystem,and real-time monitoring of vegetation is of great significance.In order to monitor the status of vegetation on the earth quickly and on a large area,remote sensing technology is used to invert multiple parameters of vegetation to understand the growth and production of vegetation and other methods are widely used.At present,there are few inversion studies on vegetation parameters in the hilly areas of northeastern Sichuan,and there are fewer related studies using Sentinel-1/2 images.In this thesis,Sentinel-1/2 images and UAV images are used for the inversion of vegetation coverage in the hilly area of northeastern Sichuan,the identification of vegetation types,and the inversion of rice height and biomass.These studies are helpful to monitor the status of vegetation in the hilly area of northeastern Sichuan,can provide a basis for more vegetation parameter research in the hilly area of northeastern Sichuan,and provide algorithmic reference for related research in other regions.On the basis of summarizing the inversion of various vegetation parameters using remote sensing technology at home and abroad,this thesis realizes the inversion of vegetation parameters in the hilly area of northeastern Sichuan.The specific research and results are as follows:(1)The hilly area of northeastern Sichuan is rich in summer vegetation types,and the vegetation coverage is in good condition,which is suitable for the research work of vegetation coverage inversion.In this thesis,cloudless Sentinel-2 images and UAV images are used for inversion and verification of vegetation coverage.Because spectral variability and mixed pixels are common in optical remote sensing images in the hilly area of northeastern Sichuan,this thesis introduces spectral normalization technology and spectral mixing analysis algorithm.Use five spectral hybrid analysis algorithms based on spectral normalization,FCLSU,PLMM,ELMM,GLMM,and Deep GUN,to invert the vegetation coverage of the cloudless Sentinel-2 image,and then use the vegetation coverage obtained by the drone image as a reference Image,compare and analyze the accuracy of inversion of vegetation coverage by five algorithms.The results show that among the five algorithms,the FCLSU,PLMM,and Deep GUN algorithms have higher inversion accuracy,and the FCLSU algorithm has the highest inversion accuracy.However,combined with field investigations,it is found that the results of Deep GUN algorithm inversion are the closest to the distribution of surface vegetation.In addition,FCLSU has the highest operating efficiency among the five algorithms.(2)Since the optical images in the hilly areas of northeastern Sichuan are susceptible to cloud and fog pollution,there are few cloud-free images available.In order to better monitor the dynamics of surface vegetation in real time,and at the same time increase the utilization rate of Sentinel-2images covered by clouds.This thesis combines Sentinel-1 to remove the clouds and fog on the Sentinel-2 image,and then uses the ELMM algorithm under spectral normalization to invert the vegetation coverage,and uses the neural network classification result as the verification data to evaluate the accuracy of the vegetation coverage inversion.The value of the inversion result R is about 0.95,and the RMSE is about 0.14.The experiment shows that the Sentinel-2 image after cloud removal has practical value and can be used for the inversion of vegetation cover,and the inversion result is close to the real situation of surface vegetation cover.(3)The hilly areas of northeastern Sichuan are rich in vegetation types,but the plots are fragmented,and the vegetation growing on adjacent plots is not continuous.Although the use of open source remote sensing images to identify large-area vegetation types is meaningful,it is full of challenges.This thesis uses cloudless Sentinel-2 images and uses Spectral-Spatial Fully Convolutional Networks(SSFCN)to classify images to realize vegetation type recognition,and finally get the total accuracy of classification.Statistics show that the average classification accuracy of each feature is greater than 90.14 %,the classification accuracy of the entire Xichong County has reached 94.36%,which shows that this classification method is suitable for feature recognition in hilly areas in northeastern Sichuan,and the recognition effect is good.In addition,the fragmented land plots in the hilly area of northeastern Sichuan,coupled with the limited spatial resolution of remote sensing images used,are currently limited in the types of vegetation that can be accurately identified.There are three types of vegetation covering large areas of rice,lotus root,and woodland.Among them,the classification of rice The accuracy reached93.06%,which shows that the rice area identified by this method has a better effect.(4)Rice has a long history of planting in Sichuan and is an important crop in Sichuan.In order to better understand the status of ground rice,real-time monitoring of rice height and biomass is of great significance.This thesis uses the SSFCN algorithm to identify the rice planting range,and combines the Sentinel-1 image to use the multiple linear regression model(MLR)and the semi-empirical physical model to invert the height and biomass of the rice in the entire study area,and compare and analyze the two algorithms.Retrieve the accuracy of rice height and biomass.In the inversion results of rice height:the RMSE and MAE of the MLR are 7 and 6,respectively;the RMSE and MAE of the semi-empirical physical model VV and VH bands are both greater than 13;in the inversion results of the biomass: the RMSE and MAE of the MLR are respectively 0.42 and 0.37;the RMSE and MAE of the VV and VH bands of the semi-empirical physical model are both 0.52 to 0.99,which shows that in the hilly area of northeastern Sichuan with complex topography,the two models are effective in retrieving rice height.However,the accuracy of inversion of the MLR model is better than that of the semi-empirical physical model;in addition,the accuracy of the inversion of rice biomass by the two models is poor and needs to be improved in the future.Among them,the accuracy of inversion of the biomass by the MLR model is slightly better.(5)Leaf area index is one of the important physical quantities that characterize the state of vegetation,and it is closely related to the height and biomass of vegetation.Based on the leaf area index,this thesis uses a multiple linear regression model and a dielectric radiation transfer model to invert rice height and biomass respectively,and compares and analyzes the accuracy of the two models in inverting rice height and biomass.In the inversion results of rice height,the RMSE and MAE of the multiple linear regression model are 11 and 9,respectively,and the RMSE and MAE of the radiation transmission model are about 7.254 and 6.083,respectively.In the inversion results of biomass: the RMSE and MAE of the multiple linear regression model are 0.53 and 0.43,respectively,and the RMSE and MAE of the radiative transmission model are 0.77 and 0.67,respectively,which shows that: in the hills of northeastern Sichuan with complex terrain and fragmented land In the region,the rice height inversion effect of the two models is better,which is closer to the true height of the surface rice,and the radiation transfer model has better inversion accuracy;in addition,the accuracy of the rice biomass inversion by the two models is slightly worse,and the follow-up is needed Continue to improve,and the inversion accuracy of the multiple linear regression model is better than that of the radiative transfer model. |
PDF Full Text Request