Cross-Continental Bayesian InSAR Forest Height Estimation

Abstract

The detection and long-term analysis of forest disturbances is a milestone in large-scale forest monitoring, climate-change mitigation strategies and biodiversity preservation (FAO, 2020). This task can be addressed through repeated estimates of key forest attributes, such as canopy height, whose temporal evolution captures forest dynamics and enables the interpretation of ecosystem changes. In this context, uncertainty estimation becomes as critical as the estimation of the forest observables themselves, since change detection can only be deemed reliable when the observed variations exceed the associated uncertainty.

Deep learning-based approaches applied to Interferometric SAR (InSAR) data have recently demonstrated state-of-the-art performance in forest height estimation at national scales (Carcereri et al., 2024). However, despite the large spatial extent of the study, both training and testing remain geographically confined to Gabon (Central Africa), limiting confidence in the network deployability at broader scales. In this work, to assess cross-domain generalization, we design a comprehensive experimental setup spanning two geographically distant tropical regions: Gabon and French Guiana (South America). Both regions are largely covered by primary tropical rainforest and include mangrove ecosystems along the Atlantic coastline. French Guiana is approximately 98% forest-covered within the Guiana Shield and exhibits a continuous mangrove belt, whereas the Gabonese coast is more heterogeneous, with mangroves occurring in scattered areas. This configuration provides a stringent cross-continental test case, moving beyond the single-country validation commonly adopted in the literature (Carcereri et al., 2024; Becker et al., 2023; Mahesh and Hänsch, 2024).

Building on the Bayesian formulation proposed in Ghio (2025), we extend the deterministic approach of Carcereri et al. (2024) by explicitly modeling predictive uncertainty as an aleatoric component, that is, the irreducible component arising from the data-generating process, and an epistemic component, that is, the residual component arising from multiple plausible parameter sets that fit the data equally well. Here, epistemic uncertainty is approximated via Bayesian model averaging using a deep ensemble of five independently trained networks, and the two components are combined under the law of total variance to yield a robust uncertainty estimate attached to each forest-height prediction, directly supporting monitoring-oriented applications.

Reference canopy heights are derived from NASA LVIS measurements (NASA Goddard Space Flight Center, 2024): for Gabon, we use data acquired during the 2016 AfriSAR campaign, while for French Guiana we rely on airborne surveys over dense Amazon rainforest in 2021. Interferometric observables are extracted from a large multi-temporal archive (2010-2024) of approximately 1000 TanDEM-X bistatic acquisitions, processed at 25 m spatial resolution.

The experimental analysis includes in-country baselines, cross-continental transfers with and without domain adaptation via fine-tuning, and joint training across continents. Results show well-calibrated uncertainty estimates and state-of-the-art in-country performance consistent with Carcereri et al. (2024) and Qi et al. (2025), with the additional benefit of the ensemble yielding a measurable improvement in regression performance; under cross-continental transfer, a meaningful relationship between predicted and reference heights is preserved in both directions, supporting the domain-generalization capability of the Bayesian framework. Domain adaptation yields mixed outcomes: it provides a modest improvement under transfer, yet exhibits signs of catastrophic forgetting when re-evaluated on the source domain. Joint training proves the most effective strategy, as it broadens the training distribution by exposing the networks to heterogeneous forest conditions and acquisition geometries, thereby emerging as the most promising pathway toward pan-tropical scalability.

Ultimately, this study demonstrates that InSAR-driven Bayesian deep learning for forest-parameter retrieval can be extended beyond single-country settings to intercontinental scenarios, laying the groundwork for globally reliable forest-monitoring products. These findings are directly relevant in the context of current and upcoming European SAR missions, such as Sentinel-1, Biomass, ROSE-L and Harmony.