STAMI: A machine learning-based small-scale tomography-aided multibaseline (Pol)InSAR forest height inversion framework

• TomoSAR vertical structure enhances multibaseline (Pol)InSAR forest height mapping. • Machine learning is used to reduce data burden and improve mapping performance. • Small-scale TomoSAR assists in obtaining large-scale vertical structure knowledge. • Reweighted estimation retains k z that best matches the forest vertical structure. • Provide guidance for multibaseline (Pol)InSAR mission design and data acquisitions. The complementary measurements of model-based multibaseline (polarimetric) synthetic aperture radar interferometry ((Pol)InSAR) and data-driven SAR tomography (TomoSAR) have become the development trend of forest height mapping missions. Their common core is the characterization of forest vertical structure, which promotes the introduction of TomoSAR reconstructed vertical structure function into (Pol)InSAR forest height inversion model. However, to provide knowledge of vertical structure, TomoSAR measurements have to cover the entire forested area. The high data acquisition cost and multibaseline observation errors hinder its application in large-scale wall-to-wall forest height mapping. Therefore, based on machine learning, this study proposes a small-scale tomography-aided multibaseline (Pol)InSAR (STAMI) framework to derive vertical structure knowledge at low data burden and estimate forest height with high accuracy. The synergy inversion is reformulated as the cross-scale clustering and classification task of the vertical structure knowledge. We conduct TomoSAR only in a small-scale forested area, and establish the mapping relationship between interferometric features and vertical structure with spectral clustering and Support Vector Machine (SVM). In large-scale mapping area, we obtain vertical structure information with only three-baseline (Pol)InSAR, which is used to construct different forest height inversion models. Finally, a reweighted method is applied to resist observation errors to different vertical structures for improving accuracy of three-baseline inversion. Airborne L- and P-band single and fully polarimetric datasets covering boreal and tropical forests are used to demonstrate the superiority of the proposed framework. This research is valuable for data acquisition planning and algorithm application of multibaseline (Pol)InSAR missions aimed at global forest mapping.