Paving the Way for Uncertainty-Aware Forestry Products Using TanDEM-X and AI
TL;DR
A novel Bayesian DL framework for the uncertainty-aware estimation of forest height from single-pass TanDEM-X InSAR data, marking its first application to synthetic aperture radar interferometry and paving the way for real-world prediction of uncertainty-aware forest canopy height products from InSAR data.
Abstract
As forests face increasing anthropogenic pressure, the robust, large-scale remote sensing of key biophysical parameters, such as forest height and above-ground biomass density, has become essential to guide conservation efforts and climatological analyses. While deep learning (DL) applied to interferometric synthetic aperture radar (InSAR) data has achieved state-of-the-art performance, its operational potential for long-term forest monitoring remains limited by a lack of uncertainty reporting, which is essential for error propagation, and for evaluating prediction reliability and temporal stability. This study introduces a novel Bayesian DL framework for the uncertainty-aware estimation of forest height from single-pass TanDEM-X InSAR data, marking its first application to synthetic aperture radar interferometry. The approach is validated through a case study in Norway, using national airborne laser scanning data to assess the calibration of the self-reported uncertainties, and to examine the effects of tree species variability and temporal mismatches in training pairs on model generalization. To capture total predictive uncertainty, the framework explicitly models the aleatoric component during estimation while integrating epistemic uncertainty through the comparison of different intra- and interbasin approximation strategies. In addition, the robustness of the model is evaluated under out-of-distribution (OOD) conditions (i.e., data domains absent from the training set), reflecting the challenges encountered in operational remote sensing. The results demonstrate robust generalization performance and the generation of well-calibrated uncertainty maps under in-distribution conditions, while highlighting the critical necessity of complementing data-driven models with an ad hoc OOD detector to reliably manage OOD scenarios. The resulting framework paves the way for real-world prediction of uncertainty-aware forest canopy height products from InSAR data.