Amazon Wetlands Map 2025: a 25 m, five-class, multi-season (2014-2025) wetland map of Amazonia latissimo sensu (V1.0)

The Amazon Wetland Map V1 is a 25 m, five-class wetland classification covering a 3.5 million km2 focus area spanning the mainstem of the Amazon River and its major tributaries. This first release (V1.0) comprises a classification of the core Amazon River mainstem covering an area of 974,500 km^2. Future releases will cover the entire area of Amazonia latissimo sensu. The map is delivered as a single Cloud-Optimized GeoTIFF in the South-America Albers equal-area projection (ESRI:102033) at 25 m, with two bands: - Band 1 - classes (uint8): 0 - nodata, 1 - water, 2 - herbaceous wetland, 3 - woody wetland, 4 - upland, 5 - built. - Band 2 - probability (uint8): the classifier's maximum class probability, stored as 0-200 (divide by 200 for a 0-1 confidence). Use it to mask low-confidence pixels. How it was made (summary): Seasonal (wet/dry) composites of ALOS-2 PALSAR-2 L-band and Sentinel-1 C-band radar, Harmonized Landsat-Sentinel optical vegetation/moisture indices, Google Satellite Embeddings, and terrain descriptors were generated in Google Earth Engine. Reference samples were built with a human-in-the-loop active-learning workflow and a gradient-boosted decision-tree classifier (LightGBM) was applied basin-wide, followed by a small-patch sieve and a built-up burn-in from Google Dynamic World. Full methods will be released in an upcoming companion data paper (see Related works). Version 1 accuracy.This is intended as the first release of a living map that will be continuously updated and improved. Accuracy reported for this vesion is preliminary (internal cross-validation and a small set of reference areas); a forthcoming V1.1 will add a design-based accuracy assessment and an open community sampling workflow. How to use. Open amazonia_wetlands_map_v1.0.tif in QGIS or any GDAL-based tool; the internal overviews give fast multi-scale display. Apply the class legend above; use band 2 as a confidence layer. Citation: Please cite both this dataset (DOI above) and the companion data paper when avaiable (see Related works). KNOWN ISSUES (V1.0) Accuracy is PRELIMINARY: internal cross-validation plus a small set of purposively chosen reference areas - NOT a design-based probability sample - so no bias-corrected, basin-wide accuracy or area estimates are reported. A design-based assessment (Olofsson et al. 2014) is planned for V1.1. Reference labels came from an active-learning loop with a confidence filter: labels are pure but their selection is not statistically independent, so cross-validation scores are optimistic relative to true map accuracy. The mosaic mixes two slightly different terrain-normalisation settings used during development; the effect on class labels is expected to be minor but is not yet quantified (a uniform re-normalisation is planned for V1.1). Some class pairs are intrinsically hard to separate (e.g. herbaceous wetland vs seasonally dry herbaceous upland); local errors are expected there. The built class (5) is inherited wholesale from Google Dynamic World v1 and shares its errors. Seasonal-window and tile-edge artefacts may occur where input coverage is sparse.