Dataset

1. Overview

In accordance with the ALOS-2 Basic Observation Scenario, PALSAR-2 ScanSAR data are collected across the Pan-tropical zones (South America, Central Africa, Southeast Asia) every 42 days, resulting in 9 observations per year over each region.
This dataset maps inundation extent and changes over time in forested wetland ecosystems across northern South America and Central Africa from January 2015 to December 2025, derived from time series of PALSAR-2 ScanSAR data. In South America the maps cover the Amazon, Orinoco and Guiana Shield river basins, the Pantanal and river basins in Colombia, Ecuador and Peru; in Central Africa they cover the central Congo Basin, Ogooué, and Nyong rivers.

2. Dataset description

3. Methodology

3.1 SAR pre-processing

Input data: ALOS-2 PALSAR-2 ScanSAR L-band SAR data, processed by JAXA to conform with the CEOS Analysis Ready Data (CEOS-ARD) format specifications for Normalized Radar Backscatter (CEOS-ARD SAR v1.0) where the data have been subject to ortho-rectification and Radiometric Terrain Correction (RTC), and are provided as Gamma-0 backscatter (Image credit: © JAXA/EORC, 2026).

Using the CEOS Analytics Lab cloud server, the ScanSAR products were speckle-filtered (Lee, 5x5), reprojected from UTM to geographic coordinates and resampled to 1.6 arcsec (~50 m) pixel spacing. Additional radiometric corrections for antenna pattern and scene-to-scene backscatter variations were applied to mitigate residual radiometric variations. The South America region of interest (12.5 M㎢) was covered by 220 path/row frames and 21,000 individual ScanSAR scenes for the Sept 2014 - Aug 2026 time period (ALOS-2 cycles 004-315). The Central Africa region (5.6 M㎢) was covered by 95 frames and about 8,000 scenes.

3.2 Classification

Based on a decision tree classification approach, the workflow was developed to handle the large time-series stacks of data involved. Each ScanSAR scene available was classified with an incidence-angle dependent threshold decision tree, including thresholds for relative backscatter changes in the time stack, into 5 classes: inundated vegetation (IV), open water (OW), submerged vegetation (SV), exposed river banks (MB) and other land cover (Other LC). Non-forest areas were masked using a combined mask generated from the JAXA FNF map, ESA WorldCover, MapBiomas, INPE TerraClass, GFW Tropical Tree Cover the ALOS-2 2014-2017 maximum inundation extent (Rosenqvist J. et al., 2020). The OW, SV and MB classes were confined to the JRC Global Surface Water dataset, Max Extent 1984-2021. Areas with a slope ≥10 degrees were masked using a terrain mask derived from the ASF GLO-30 HAND dataset.

The average acquisition success rate for ALOS-2 ScanSAR over South America and Central Africa was found to be about 82%, resulting in sporadic gaps in the basin-wide coverages. A large share of the acquisition gaps were self-mitigated thanks to the extensive overlap between adjacent ScanSAR passes and the actual loss of data over the regions of interest was estimated at 7.6%. Such remaining pixels with no acquisition in a given observation cycle were gap-filled through temporal filtering over the observation cycles immediately before and after. When indecisive, or when no adjacent cycle data were available, gap-filling was undertaken through a hydrological model approach.

The Inundation Duration maps provided here were derived from per-pixel counts from the wetland classifications described above, with the pixel value indicating the number of observations in which a pixel was classified as IV during each year.

3.3 Accuracy assessment

Classification accuracy was assessed against 6,863 visually interpreted reference points distributed across the South America region. The overall accuracy was estimated at 91.3%, with a commission error of 2.9% and an omission error of 41.5%. With the dominant error being under-detection of inundation, extent and duration should be regarded as conservative (minimum) estimates.

The correlation between inundation extent and river stage was assessed at 38 river gauges in the Amazon Basin. Mapped inundated area followed river stage with a median R² of 0.88 for fully imaged observation cycles and 0.83 for gap-filled cycles, indicating the robustness of the gap-filling approach.