Speaker
Description
In the realm of artificial intelligence, specifically utilizing methodologies such as machine learning and deep learning, a conspicuous display of substantial potential across various parameter estimation problems has been demonstrated. However, such AI techniques are often employed without the incorporation of domain-specific knowledge or expertise raising concerns about the explainability and robustness of the implemented methodologies.
In contrast, physical models (PMs) offer a significantly enhanced level of deterministic robustness. However, it is imperative to recognize that these models can exhibit performance limitations owing to their inherent simplicity and/or strictness. Moreover, the accuracy of their inversion process is circumscribed by the assumptions and simplifications that underlie them, particularly those applied to the vertical reflectivity function, which are prerequisites for achieving a well-balanced inversion problem.
As a result, it becomes imperative to advocate for hybrid modeling approach by the integration of AI techniques with physical models, especially in the context of forest height estimation derived from TanDEM-X coherence measurements.
Accurate estimation of forest height is crucial for understanding forest structure and biomass, which in turn plays a pivotal role in climate change mitigation and ecosystem management. In this study, we propose a novel hybrid modeling approach that combines machine learning techniques and physical models to invert forest height from TanDEM-X InSAR (Interferometric Synthetic Aperture Radar) data. This approach might be relevant for the Biomass mission for understating the forest and its structures.
The conventional methods for forest height estimation from InSAR data often rely solely on physical models, which are based on Polarimetric InSAR (Pol-InSAR) measurements as an established application demonstrated and validated at large scales for a wide variety of boreal and tropical forest sites at different frequencies (from X- down to P- band) [1], [2]. However, it may suffer from limitations in complex forest environments with varying topography forest types and underlying assumptions [3]. On the other side, machine learning approaches have demonstrated great potential in capturing complex relationships within data but may lack the physical interpretability required for robust forest height estimation. Furthermore, the limited availability of training datasets that covers a variety of condition might cause issues with the generalizability of the machine learning model.
The hybrid modeling approach bridges this gap by integrating the strengths of both approaches. We leverage a dataset of TanDEM-X InSAR observations, and LVIS forest height measurements over Gabon, Africa. The MultiLayer Perceptron (MLP) is employed to learn complex patterns and relationships from the data, to derive the vertical reflectivity profile. Whereas the vertical reflectivity profile is defined as a set of coefficients utilizing a Legendre series decomposition [4].
We incorporate physical models to enhance the interpretability and generalizability of the predictions. By fusing the machine learning predictions with the physical model outputs, we obtain a hybrid forest height estimation that combines the advantages of both approaches. Moreover, it provides optimum performance and robustness. We incorporate multi-model data for the forest height estimation from single baseline single-polarimetric TanDEM-X interferometric coherence measurements. We discuss the challenges that need to be addressed and how to integrate a multi-modal (LiDAR, multi-spectral images, Polarimetric SAR, etc.) in a general hybrid modelling framework.
Keywords: Forest height inversion, TanDEM-X InSAR data, hybrid modeling, machine learning, physical models, remote sensing.
[1] V. Carcarra-Bes, M. Pardini, C. Choi, R. Guliaev, and K. P. Papathanassiou, “Tandem-X and Gedi Data Fusion for a Continuous Forest Height Mapping at Large Scales,” pp. 796–799, 2021, doi: 10.1109/igarss47720.2021.9554655.
[2] S. K. Lee et al., “Spaceborne GEDI and TanDEM-X data fusion for enhanced forest height and biomass,” AGUFM, vol. 2018, pp. B44E-14, 2018, Accessed: May 03, 2021. [Online]. Available: https://ui.adsabs.harvard.edu/abs/2018AGUFM.B44E..14L/abstract
[3] A. T. Caicoya, F. Kugler, K. Papathanassiou, P. Biber, and H. Pretzsch, “Biomass estimation as a function of vertical forest structure and forest height - Potential and limitations for Radar Remote Sensing,” in 8th European Conference on Synthetic Aperture Radar, 2010, pp. 1–4.
[4] S. R. Cloude, “Polarization coherence tomography,” Radio Sci, vol. 41, no. 4, Jul. 2006, doi: 10.1029/2005RS003436.
