Speaker
Description
Accurate characterization and continuous monitoring of forest canopy height are essential to guide conservation efforts, evaluate temporal stability, and support climatological analyses under increasing anthropogenic pressures. While deep learning (DL) models applied to interferometric synthetic aperture radar (InSAR) data have been shown to achieve state-of-the-art performance, their operational utility has been hindered by a lack of uncertainty reporting, which is critical for error propagation and determining prediction reliability. This work introduces a novel Bayesian deep learning (BDL) framework for the uncertainty-aware estimation of forest canopy height using single-pass TanDEM-X InSAR data, marking its first application to SAR interferometry. The framework explicitly models aleatoric uncertainty through a heteroscedastic neural network and captures epistemic uncertainty by evaluating multiple intra- and inter-basin approximation strategies, including Stochastic Weight Averaging (SWAG), Deep Ensembles, and Multi-SWAG. The proposed approach was evaluated across a country-scale dataset in Norway with national airborne laser scanning (ALS) data as a reference, and demonstrated excellent in-distribution performance. All configurations yielded accurate forest height estimates, with the best achieving a bias of 0.01 m, a mean absolute error (MAE) of 2.38 m, a root mean squared error (RMSE) of 3.17 m, and a coefficient of determination (R2) of 0.76. Furthermore, all models provided well-calibrated and consistent uncertainty estimates under in-distribution conditions, with in particular the inter-basin approaches achieving a total predicted error (TPE) of 99% and an expected calibration error (ECE) of 0.047. However, when tested under out-of-distribution (OOD) conditions, which we emulated by systematically occluding products acquired with different height of ambiguity values, all Bayesian models were shown to significantly underestimate prediction errors. To resolve this, we coupled the BDL model with an ad hoc OOD detector, which successfully flagged domain-shifted samples, restoring uncertainty calibration and regression performance to in-distribution levels. This framework establishes a robust, reliable, and operational paradigm for generating the next generation of uncertainty-aware forest parameter maps from spaceborne InSAR data.