Research On The Inversion Model Of Vertical Structure Of Vegetation Based On Pol-InSAR

As a major ecological community on Earth,vegetation is an extremely important component of terrestrial ecosystems and plays a dominant role in the global carbon and water cycle.Vertical structure of vegetation is an important issue in ecosystem research,and polarized interferometric synthetic aperture radar is a powerful technique for extracting forest parameters by inversion of physical scattering models.The study area of this paper was selected as Liupan Mountain in Guyuan City,Ningxia Autonomous Region,which is rich in vegetation types and has complex and variable slopes because it is a mountain range,so it is a good choice area for the application of Pol-InSAR technology.The classical three-stage inversion algorithm based on the RVoG model suffers from underestimation,for which an improved methodological study of the three-stage inversion algorithm based on the S-RVoG model is carried out in this paper.Considering that simplified scattering models limit the inversion accuracy unless multiple baseline measurements are used,which would lead to increased costs and computational difficulties,a complex-valued convolutional neural network combined with Vision Transformer inversion method was investigated to obtain higher accuracy of vegetation vertical structure inversion.The main work content is as follows:(1)Improved three-stage inversion algorithm for the S-RVoG model with linear extinction coefficients.An improved three-stage inversion algorithm based on the S-RVoG model with linear extinction coefficients was introduced to account for the underestimation phenomenon(r=0.686,RMSE=4.16 m)in the three-stage inversion algorithm based on the RVoG model.It was found that the inversion accuracy was improved with the introduction of the linear extinction coefficient(r=0.889,RMSE=3.23).(2)CV-CMT network model.Although the study in(1)above achieved good inversion accuracy,there is still underestimation.To address the problem of further improving the inversion accuracy,the CVCMT network model is proposed based on CV-CNN in conjunction with the Vision Transformer network model.It was found that the inversion accuracy based on CV-CMT network model was(r=0.913,RMSE=1.88),and after improvement,the inversion accuracy was improved(r=0.931,RMSE=1.79).The highest inversion accuracy was achieved by the CV-CMT-based model,which improved the correlation coefficient r by 0.18 and reduced the root mean square error(RMSE)by 0.09 compared with the CVCNN-based model.The inversion accuracy of the L-S-RVoG model and the other two traditional RVoGbased models is lower than that of the deep learning model.Among them,the three-stage reproduction algorithm based on the RVoG model is the lowest(r=0.686,RMSE=4.16),and the improved L-S-RVoG model is better than Based on the improved three-stage inversion algorithm based on RVoG model underestimation compensation,the model correlation coefficient r increased by 0.045,and the root mean square error(RMSE)decreased by 0.35.