Evaluating P- and C-band spaceborne SAR for refined tropical ecosystem mapping in Amazonia under sparse ground-truth conditions

Remote sensing plays a central role in providing qualitative and quantitative information on forest ecosystems for sustainable management and forest carbon inventory. The ESA BIOMASS satellite mission launched in 2025 introduces the first global P-band Synthetic Aperture Radar (SAR) system dedicated to forest structure and biomass, but its potential for refining tropical forest ecosystem maps remains largely unexplored. Here we compare P-band (~70cm wavelength) BIOMASS data with C-band Sentinel-1 SAR to assess their respective ability to discriminate structurally and hydrologically different forest ecosystems in Brazilian and Bolivian Amazonia, including terra firme forest, flooded forest and forest–wetland mosaics, in the presence of sparse solid ground-truth data.

We use MapBiomas Amazonia and MapBiomas Bolivia as primary land-cover references, taking advantage of their annual time series from 1985 to 2024 to address label noise. Because these maps are not error-free at the pixel level, we develop a noise-labelling pre-processing workflow to derive high-confidence forest samples at the epoch of the first BIOMASS acquisitions (2025). The workflow combines (i) spatial homogeneity constraints (distance to class boundaries, neighbourhood purity, minimum patch size), (ii) temporal stability of the MapBiomas class history (to identify pixels with persistent forest or flooded forest trajectories), and (iii) physical plausibility checks using auxiliary optical and terrain indicators. Pixels that satisfy these criteria are retained as reliable proxies for different forest ecosystem types.

For these filtered samples we extract BIOMASS Detected Ground-range Multi-looked (DGM) backscatter and Sentinel-1 Ground-range Detected (GRD) backscatter, derive polarisation ratios and simple texture metrics, and quantify within-class variability and between-class separability for both frequencies. We pay particular attention to forest–non-forest transitions and to distinctions among terra firme forest, flooded forest and adjacent forested wetlands that are relevant for high-carbon stock and peat-forming systems. Preliminary results from the Brazilian test site indicate that P-band reduces within-class variance in forested classes and enhances separability between terra firme and flooded forests compared to C-band alone, while C-band performs comparably or better for some open and anthropogenic land covers. By extending the analysis to Bolivian Amazonia and to a richer legend of forest and forest-wetland classes, and by testing a similar workflow for peatland-prone flooded forests, this study provides a first evaluation of the potentials and limitations of BIOMASS P-band SAR for tropical forest applications under sparse ground-truth conditions.