Research On Forest Height Inversion Algorithms Based On Repeat-pass Polarimetric Interferometric Synthetic Aperture Radar

As the main body of terrestrial ecosystem,forest plays an irreplaceable role in the process of human survival and development.As one of the important renewable resources,the exploration and protection of forest resources has become an important issue in the international environment.Forest height is the most important vertical structure parameter of forest,and its quantitative measurement will effectively promote the development of forestry remote sensing.Polarimetreic SAR Interferometry(Pol In SAR)holds unique advantages in forest height estimation due to its sensitivity to scattering centers at different heights.Therefore,it is of great significance to carry out research on the forest height.Based on Pol In SAR theory,this paper conducted inversion of forest height by dual-baseline?polarization coherence tomograpgy?machine learning respectively,and compared the accuracy with the traditional physical model.The main contents and results are as followed:(1)Aiming at the problem of ground scattering contribution and temporal decorrelation in single baseline P-band inversion,a dual-baseline inversion method based on RVo G+Volume Temporal Decorrelation(RVo G+VTD)model is proposed.In this method,the ground to volume amplitude ratios of the same polarization mechanism between different baselines are used as the condition to constrain two baselines.A look-up table algorithm is then used to find the corresponding height as the difference between the ground to volume amplitude ratios reaches to the smallest.In this way,the assumption of zero ground to volume amplitude ratio in single baseline can be avoided,and the forest height with less influence of ground scattering contribution can be obtained.Meanwhile,based on the temporal decorrelation model RVo G+VTD,the calculated solution is limited to the fuzzy interval along the phase line caused by the ground scattering contribution and the radial fuzzy interval of the complex coherence line between the origin and the volume.Finally,the forest height which ameliorates the effects of ground scattering contribution and temporal decorrelation can be obtained.The results show that the forest height inversion error caused by ground scattering contribution and temporal decorrelation is closely related to the vertical wavenumber6)and forest height h_v.Therefore,the simulation experiment is further used to explore how the forest height error changes under different6)andh_valong the change of ground to volume amplitude ratio and temporal decorrelation coefficient.Compared with the traditional three-stage inversion method for BL1(R~2=0.38,RMSE=6.76m)and BL2(R~2=0.45,RMSE=7.9m)and the fixed extinction coefficient method for BL1(R~2=0.46,RMSE=5.88m)and BL2(R~2=0.47,RMSE=4.42m),this method has higher inversion accuracy(R~2=0.6,RMSE=3.48m).(2)A novel forest height retrieval method based on tree species information using Polarization Coherence Tomography(PCT)is proposed to solve the error caused by the absence of vertical heterogeneity.Based on the tree species information,the method uses PCT algorithm to establish the vertical relative reflectance of the same tree specie proportion pixel,and calculates the corresponding tomographic height of its maximum relative reflectance,then calculates its power loss with the height measured by Li DAR as the benchmark,finally obtains the power loss distribution in the research area.Based on the tomographic height of a single pixel,the forest height containing tree structure information can be obtained by compensating the corresponding loss height according to the corresponding power loss.The results show that the vertical relative reflectance of the scottish pine is different from that of the norwegian spruce.The norwegian spruce has a higher phase center and a larger relative reflectance span.Compared with the traditional three-stage method(R~2=0.35,RMSE=7.09m),the accuracy of the height obtained by this method is obviously improved because the vertical structure of different tree species is taken into account(R~2=0.51,RMSE=4.96m).(3)The random forest algorithm is used to explore the potential of correlated attributes of coherence region with different baselines in L and P bands in forest height inversion.The selected training features describe the sensor and terrain,the position and direction of the coherence region,the dimension closest to the ellipse,and the similarity between the shape of the coherence region and the theoretical ellipse.Sparse Li DAR samples are used as labels to establish relationships with features,and then applied to different regions.The results obtained are explained in terms of accuracy,transferability and interpretability.The results show that compared with the inversion accuracy of physical model features,the selected features can greatly improve the accuracy.As for BL1,based on the input features of three-stage inversion method and the fixed extinction method,R~2 is improved by 0.16 and 0.13,RMSE is decreased by0.36m and 0.33m,respectively.For BL2,R~2 increased by 0.15 and 0.13,RMSE decreased by 0.35m and 0.3m,respectively.When comparing different bands,the application effect of L-band is better than that of P-band,with R~2 reaching to 0.67 and RMSE decreasing to 1.27m.A possible explanation is that the L band's coherence region shape and Canopy Height Model(CHM)has higher linear correlation and is more important than other features in machine learning modeling.