Conference Agenda

The detection and long-term analysis of forest disturbances is a milestone in large-scale forest monitoring, climate-change mitigation strategies and biodiversity preservation [1]. 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 [2]. 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 ∼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 [2], [3], [4].

Building on the Bayesian formulation proposed in [5], we extend the deterministic approach of [2] by explicitly modeling predictive uncertainty as an aleatoric component (i.e., the irreducible component arising from the data-generating process) and an epistemic component (i.e., 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 [6]: 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 ∼ 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 [2] and [7], 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.

[1] Food and Agriculture Organization of the United Nations, Global Forest Resources Assessment 2020. Rome: FAO, 2020, isbn: 9789251329740. doi: 10.4060/ca9825en.

[2] D. Carcereri, P. Rizzoli, L. Dell’Amore, J.-L. Bueso-Bello, D. Ienco, and L. Bruzzone, “Generation of country-scale canopy height maps over gabon using deep learning and tandem-x insar data,” Remote Sensing of Environment, vol. 311, p. 114 270, 2024. doi: 10.1016/j.rse.2024.114270.

[3] A. Becker, S. Russo, S. Puliti, N. Lang, K. Schindler, and J. D. Wegner, “Country-wide retrieval of forest structure from optical and sar satellite imagery with deep ensembles,” ISPRS Journal of Photogram- metry and Remote Sensing, vol. 195, pp. 269–286, 2023. doi: 10.1016/j.isprsjprs.2022.11.011.

[4] R. B. Mahesh and R. Hänsch, “Forest height estimation with tandem-x sar and insar features using deep learning,” IEEE Geoscience and Remote Sensing Letters, vol. 21, 2024, issn: 1558-0571. doi: 10.1109/lgrs.2024.3474252. [Online]. Available: http://dx.doi.org/10.1109/LGRS.2024.3474252.

[5] F. Ghio, “Deep-learning-basierte erstellung einer europäischen waldhöhenkarte aus radarinterferometrischen tandem-x daten,” M.S. thesis, Politecnico di Milano, Mar. 2025. [Online]. Available: https://elib. dlr.de/204650/.

[6] NASA Goddard Space Flight Center, Land, vegetation, and ice sensor (lvis) instrument, https:// lvis.gsfc.nasa.gov, 2024. [Online]. Available: https://lvis.gsfc.nasa.gov.

[7] W. Qi et al., “Mapping large-scale pantropical forest canopy height by integrating gedi lidar and tandem- x insar data,” Remote Sensing of Environment, vol. 318, p. 114 534, Mar. 2025. doi: 10.1016/j.rse. 2024.114534.

Generative AI tools were used for minor language revisions, in line with institutional policies on responsible use. The authors are solely responsible for the content of this work.

Mangrove canopy height is an important variable directly related to above-ground biomass estimation and serves as a key environmental indicator for evaluating carbon storage in blue carbon ecosystems. Mangroves are found in coastal areas and are strongly influenced by complex hydrological and environmental factors, including tides, salinity, and flooding. In particular, extreme weather events such as hurricanes can cause defoliation and branch breakage, resulting in rapid and substantial changes in canopy height. These structural changes directly impact biomass and carbon storage, underscoring the importance of quantitatively monitoring changes in mangrove canopy height over time. However, it is challenging to collect vegetation height data through field surveys over large areas. Although light detection and ranging systems and in-situ surveys, provide high-resolution and accurate information on vegetation vertical structure, their application is limited by high costs and restricted spatial coverage. To overcome these limitations, Synthetic Aperture Radar (SAR) remote sensing techniques have been widely used.

The Everglades wetlands in southern Florida are among the most important mangrove ecosystems globally. This region exhibits spatially diverse canopy structures and is designated as a protected area, requiring quantitative information on vegetation structure for systematic ecosystem management and restoration assessment. Furthermore, the impact of Hurricane Irma in 2017 has further highlighted the need for quantitative assessment of changes in canopy structure. Therefore, this study aimed to estimate mangrove canopy height in the southern part of Everglades National Park, a natural wetland. We utilized the X-band (~9.6 GHz) TanDEM-X pursuit monostatic Stripmap data acquired by the German Aerospace Center. The pursuit monostatic acquisition mode, with a short temporal baseline of approximately 10 seconds, minimizes temporal decorrelation and is advantageous for analyzing volume decorrelation. The TanDEM-X datasets used in this study had heights of ambiguity (HoA) ranging from 15 to 75 m, corresponding to perpendicular baselines of 62–343 m and incidence angles between 31° and 33°. HoA is defined as the height difference corresponding to a 2π interferometric phase cycle and is inversely proportional to the perpendicular baseline. We evaluated the sensitivity of vertical structure estimation under various observation geometries and analyzed the impact on canopy height retrieval.

Following the previous TanDEM-X Random Volume over Ground (RVoG) inversion approach for mangrove canopy height estimation (Feliciano et al., 2017), we adopted a similar preprocessing and inversion procedure. To estimate vegetation height relative to the ground, topographic phase removal using an external Digital Elevation Model (DEM) is typically performed. However, in coastal wetland areas, DEMs may not accurately represent the ground elevation due to tidal fluctuations and hydrological variability. Therefore, we used coherent scatterers rather than an external DEM to estimate and remove the ground phase referenced to the water surface, thereby isolating the volume decorrelation component. Because the TanDEM-X pairs were acquired quasi-simultaneously, their temporal decorrelation was assumed to be negligible (≈0). To better isolate vegetation-induced volume decorrelation, a common band filter was applied to mitigate spectral decorrelation caused by range bandwidth differences. In addition, signal-to-noise ratio (SNR) decorrelation was modeled and corrected using the noise-equivalent sigma-zero values provided with the SAR products. After accounting for geometric, spectral, and SNR decorrelation components, the remaining coherence was attributed primarily to volume decorrelation. Canopy height was estimated by incorporating the separated volume decorrelation and vertical wavenumber derived from the acquisition geometry into the RVoG model. To extend this framework, multi-baseline volume decorrelations and their corresponding vertical wavenumbers were jointly integrated into a unified RVoG inversion scheme, allowing the retrieval of a single optimal canopy height that minimizes baseline-dependent bias and vertical ambiguity.

As a result, when a single baseline dataset was used, the retrieved canopy heights were distributed within the 0–16 m range, with slight deviations observed near the corresponding HoA limits. These deviations reflect residual vertical ambiguity. The coefficient of determination (R²) ranged from 0.73 to 0.86, indicating that canopy height accuracy varied across HoA conditions. When all TanDEM-X pairs with different HoAs were used simultaneously, the canopy heights remained consistently within the 0–16 m range, with less dispersion at the HoA limits. In addition, R² improved to above 0.89. This result suggests that multiple HoA conditions reduce vertical ambiguity and mitigate bias associated with individual baseline configurations. Differences between the estimated and validation data are likely related to variations in vegetation density and scattering structure. Mangrove ecosystems exhibit strong spatial heterogeneity, including mixed mangrove forests, mangrove scrubs, shrublands, and wetland-forest transition zones. These diverse vegetation types, along with varying canopy densities and heights, can result in significant spatial variability in the scattering mechanism and vertical structure. Nevertheless, the estimated canopy heights showed strong agreement with global validation datasets, demonstrating the feasibility of mangrove canopy height retrieval under varying acquisition conditions.

This study demonstrates that mangrove canopy height can be estimated without an external DEM using single-polarized multi-baseline TanDEM-X data and a simplified RVoG model based on volume decorrelation. The proposed approach provides valuable information for blue carbon stock assessment and climate change research. Future integration with new sensors such as L-band NISAR and P-band BIOMASS missions may further enhance sensitivity to the ground contributions beneath mangrove vegetation and multiple canopy layers through their deeper penetration characteristics, potentially improving the accuracy of the global mangrove structure mapping.

Accurate and spatially continuous forest height information is essential for estimating aboveground biomass, carbon stocks, and ecosystem functioning, yet remains difficult to obtain in boreal and temperate regions. Spaceborne LiDAR missions such as GEDI and ICESat-2 provide high-accuracy canopy height measurements but are limited by sparse footprint sampling. Tomographic synthetic aperture radar (TomoSAR) offers a complementary solution for wall-to-wall forest height mapping. However, conventional monostatic repeat-pass configurations are strongly affected by temporal decorrelation and atmospheric phase delay variation. Here we exploit multi-static TomoSAR observations from Hongtu-1, the world’s first spaceborne multi-static SAR constellation, to retrieve forest height across boreal and temperate forests in China, The United States, and Europe. Forest vertical scattering profiles are reconstructed using a MUSIC-based TomoSAR inversion under a two-layer canopy assumption, improving vertical resolution relative to conventional Fourier/Capon beamforming. A tomographic spectral entropy criterion is applied to discriminate forested from non-forested areas. Radar-derived scattering heights are converted to true vertical forest heights by accounting for terrain slope and radar viewing geometry, with absolute calibration using airborne LiDAR where available. For sites without airborne reference data, ascending–descending orbit cross-validation is employed to assess robustness and geometry independence.

The main contributions of this study are summarized as follows:

(1) We present a comprehensive multi-site evaluation of multi-static X-band TomoSAR forest height mapping using the Hongtu-1 satellite constellation across boreal and temperate forests, covering diverse forest structures, terrains, and acquisition geometries, representing one of the first systematic assessments of spaceborne multi-static TomoSAR for forest applications beyond local feasibility demonstrations.

(2) A MUSIC-based tomographic framework is adopted to improve vertical resolution under limited baseline configurations, together with a two-layer assumption to enhance the robustness of ground and canopy top detection. The proposed approach enables reliable forest height retrieval, mitigating the vertical resolution limitations inherent to conventional Fourier/Capon-based TomoSAR methods.

(3) We quantitatively analyzed the systematic height offset between TomoSAR-derived radar phase centers and canopy tops using airborne LiDAR and GEDI observations, and investigate its variability across different forest types, providing new physical insights into the interpretation of X-band TomoSAR forest height measurements.

Through extensive validation using airborne LiDAR, GEDI observations, and ascending–descending orbit cross-comparisons, we demonstrate that multi-static TomoSAR at X-band delivers improved accuracy and substantially enhanced spatial completeness compared with spaceborne LiDAR, highlighting its complementary role in large-scale forest structure monitoring where LiDAR sampling is sparse.

Within the Sentinel User Preparation- Synthetic Aperture Radar (SUPSAR) programme, this study addresses a fundamental and timely challenge: how to retrieve forest height robustly from the near-zero baseline, repeat-pass interferometric configurations that will dominate the coming Copernicus SAR System-of-Systems. Sentinel-1, the upcoming ROSE-L mission, as well as NISAR, all operate with tightly controlled orbital tubes and regular revisit cycles, providing dense temporal sampling but limited spatial baselines. While this configuration enables systematic global monitoring, it renders classical geometric PolInSAR approaches ill-conditioned, as volume decorrelation becomes weak and height sensitivity diminishes.

Given the operational reality of 6–12 day repeat cycles and small perpendicular baselines, there is an urgent need for simple, stable, and physically consistent methods tailored to zero- or near-zero-baseline InSAR. The method should exploit temporal decorrelation behavior rather than geometric decorrelation, operate reliably under dual-polarimetric constraints, and remain extensible to multi-frequency configurations. Developing these frameworks is critical for ensuring that the Sentinel-1 and ROSE-L System-of-Systems can deliver continuous forest structure monitoring at continental to global scales.

We propose a reduced-order coherence modelling strategy for forest height estimation based upon a parameter reduction of the Random Motion over Ground (RMoG) temporal decorrelation model. Instead of directly inverting the full nonlinear scattering formulation, which contains multiple interacting physical parameters and leads to an underdetermined problem under limited observability, we reformulate the interferometric coherence as a compact height-dependent polynomial representation. Through a systematic series expansion and parameter aggregation, the complex set of attenuation, motion variance, and ground-to-volume terms is collapsed into two effective coefficients that preserve the physical behavior of the model while dramatically reducing inversion dimensionality.

This reduced parameterization transforms forest height retrieval under small-baseline conditions into a well-posed, numerically stable estimation problem. The framework is designed to operate on time-series stacks, where multiple repeat-pass interferometric pairs are jointly exploited to enhance robustness and mitigate noise. Height can then be retrieved through constrained numerical optimization, enabling consistent estimation even under dual-polarimetric acquisition modes.

In addition to deterministic inversion, we introduce a physics-guided deep learning extension in which neural networks are trained to estimate the reduced-order coefficients rather than height directly. This preserves physical interpretability while improving robustness to noise, seasonal variability, and modelling approximations. The approach remains analytically constrained, avoiding black-box regression and ensuring compatibility with multi-temporal and multi-frequency observations.

The reduced-order framework is inherently extensible. It supports single-frequency (C- or L-band) operation and can be naturally extended to dual-frequency synergy, where shared canopy height is estimated jointly from C- and L-band time-series. The formulation also accommodates multi-polarization observables and enables systematic assessment of orbit phasing strategies, temporal baselines, and frequency combinations in line with SUPSAR objectives.

By converting small-baseline interferometric coherence into a reduced, physically interpretable height model, this work opens a practical pathway for operational forest height monitoring within the Sentinel-1 and ROSE-L System-of-Systems. The approach advances state of the art beyond classical geometric PolInSAR, aligns with the high temporal density of current and future SAR missions, and establishes a scalable framework for synergistic multi-frequency forest monitoring in preparation for the next generation of Copernicus SAR capabilities

Forest stores most of the terrestrial carbon and plays a central role in climate mitigation, yet large‑scale monitoring of forest growth and vegetation water dynamics remain challenging due to limited and unscalable in‑situ measurement methods. Microwave remote sensing offers global coverage for possible above‑ground biomass estimation and vegetation monitoring, but conventional coherence‑based approaches are often affected by temporal decorrelation in dense canopies.

Therefore, this work presents a novel approach, based on Incoherent Cross‑Correlation (ICC), also known as pixel offset tracking, applied to multi‑year, Synthetic Aperture Radar (SAR) data stacks detecting seasonal and long‑term range shifts in forested regions. Centimeter‑level accuracy is achieved through rigorous model‑based and data‑driven co‑registration, followed removal of tectonic motion using GNSS‑derived velocities and temporal filtering to isolate consistent oscillatory signals. The method is also evaluated across multiple frequency bands, including C-band Sentinel-1, X-band TerraSAR-X, and L-band ALOS-2.

Over temperate forests, C-band range shifts show a consistent annual oscillation with peak‑to‑peak amplitudes of approximately 20–80 cm, superimposed on a long‑term trend of several centimeters per year. The oscillatory component shows synchronized timing within climatic zones, with peaks around March–April during spring leaf‑out and increased canopy water content. Spatial maps of sinusoid amplitude and trend demonstrate that this signal is confined to forested pixels, while urban and non‑vegetated areas show negligible range shifts. Differences between ascending and descending Sentinel‑1 orbits reveal larger amplitudes for morning descending passes, which is consistent with higher vegetation water content. Also, VH backscatter, which is more sensitive to volume scattering from canopy elements such as leaves and small branches, yields larger oscillatory amplitudes than VV.​ Correlation analyses with external datasets, including land cover maps and GEDI canopy height, demonstrates that oscillation amplitude and long-term trends scale with forest density and height.

These results demonstrate that ICC‑derived SAR range shifts provide a new observation for monitoring seasonal forest phenology and slow structural growth using intensity-only information and avoiding coherence‑related limitations of classical Interferometric SAR (InSAR). This opens a pathway to exploit existing and future SAR data for retrospective and operational analysis of forest growth, vegetation water content variability, and their links to drought stress.