An Improved Sparse Bayesian Learning SAR Tomography Method and its Application for Forest Vertical Structure Inversion

Synthetic aperture radar tomography (TomoSAR) is widely used in reconstructing forest vertical structure, but accurately locating both ground and canopy scatterers in dense forest areas remains challenging. In this paper, a novel sparse Bayesian learning (SBL) based TomoSAR method is proposed to achieve super-resolution reconstruction of forest vertical structure. Two important improvements are considered in the process of SBL SAR tomography. First, a hybrid sparse basis is employed to accurately transform and reconstruct the forest vertical structure profile with different scattering mechanisms. Second, an adaptive singular value decomposition (SVD) method is employed to address the instability issue caused by illconditioned inversion in Bayesian inference. Furthermore, leveraging high-resolution TomoSAR profiles significantly enhances the performance of forest vertical structure parameter inversion. The effectiveness of the proposed method is validated using multi-baseline P-band airborne SAR images acquired in tropical forests at two distinct test sites. The results demonstrate that the proposed method achieved high-resolution SAR tomography imaging outcomes even within a limited baseline span. In terms of forest structure parameter inversion, the root mean square error (RMSE) of inverted forest height is 2.58 m and 4.16 m compared to LiDAR measurements, while the RMSE of inverted underlying topography is 1.77 m and 5.49 m. The proposed method is instrumental in retrieving large-scale forest structure parameters, particularly in preparation for the upcoming BIOMASS mission.