Conference Agenda

The increasing availability of global Sentinel-1 observations offers new opportunities for large-scale forest structural mapping. Although temporal decorrelation in repeat-pass Sentinel-1 acquisitions limits the applicability of Polarimetric-Interferometric (PolInSAR) techniques for forest height estimation, recent studies [1] have demonstrated the potential of Sentinel-1 coherence for biophysical parameter retrieval. Such approaches rely on dense or seasonal time-series data, where the ground and volume scattering components are inferred indirectly from backscatter statistics under simplified, regionally constant assumptions.

In this study, we investigate an alternative data-driven approach for forest height estimation from Sentinel-1 interferometric observables, leveraging deep learning models trained on globally distributed SAR–LiDAR data. We assembled an extensive dataset of over 900 Sentinel-1 interferometric pairs acquired between 2019 and 2024, covering diverse forest biomes worldwide. Each pair corresponds to a single-baseline configuration with a 12-day temporal baseline and dual-polarization (VV, VH) channels. Reference forest heights were derived from GEDI L2A and ICESat-2 ATL08 products, filtered by confidence and terrain criteria, and projected into SAR geometry (azimuth and slant range). This dataset represents a uniquely comprehensive resource, encompassing a wide range of forest types, forest heights, and ecological regions across multiple continents. The Sentinel-1 interferometric pairs span diverse spatial baselines as well as varying heights of ambiguity (HoA), acquired under different InSAR geometry. This diversity of the data enables the development and evaluation of machine learning models that are robust and generalizable across biomes, acquisition geometries, and environmental conditions. To achieve this objective, a deep convolutional network was developed to map elements of the polarimetric interferometric covariance matrix to forest height. Preliminary results indicate that the model successfully learns meaningful relationships between Sentinel-1 observables and LiDAR-derived forest heights, demonstrating the feasibility of data-driven forest height mapping at the global scale. Despite the effects of temporal decorrelation, Sentinel-1 interferometric observables retain measurable sensitivity to vegetation structure: coherence generally decreases with increasing forest height, although the relationship remains nonlinear and spatially variable. Unlike TanDEM-X data, where simplified RVoG formulations can directly relate coherence to height, Sentinel-1 coherence requires alternative modeling strategies that can capture its more complex behavior.

This work contributes to ongoing efforts to integrate Sentinel-1, GEDI, and forthcoming L-band missions within the ESA’s SUPSAR framework. The research is funded by the ESA under the Sentinel2Height initiative. This initiative aims to develop data-driven approaches for global forest height mapping. Detailed results, model architecture, and assessments will be presented at the conference.

Reference

[1] Lavalle, Marco, C. Telli, Nazzareno Pierdicca, Unmesh Khati, Oliver Cartus, and Josef Kellndorfer. "Model-based retrieval of forest parameters from Sentinel-1 coherence and backscatter time series." IEEE Geoscience and Remote Sensing Letters 20 (2023): 1-5.