Reservoir maps from "Mapping the evolution of small reservoirs in Brazil from 1984 to 2025 using deep learning"

Annual maps of small (< 1 km²), stream-fed reservoirs across Brazil based on Landsat 5, 7, 8, and 9 composites. In years with overlapping sensor availability, files are provided for both sensors (e.g. ls5_2005_cloudfilt_v3.tif and ls7_2005_cloudfilt_v3.tif). The data are provided in three formats: Csvs of reservoir centroids, including surface area of each reservoir. Provided in a .zip file for all years and satellites (ls_all_reservoirs_csvs.zip). Columns: 'id_for_year_sat': Index of reservoir within map for specific year-satellite combination (e.g. ls5, 1984). 0 to total reservoir count. 'year': Year of map 'satellite': ls5, ls7, ls8, or ls9 for Landsats 5, 7, 8, and 9 'longitude', 'latitude': Coordinates of reservoir centroid. 'x_aea', 'y_aea': Reservoir centroid location in South American Albers Equal Area projection. 'area_ha', 'area_km2': Reservoir surface area in ha and km². Parquet of all reservoir centroids. Same columns as (1), but in a format that can be read in its entirety or filtered at read time. GeoTIFFs with raw model scores (0-254, NoDataValue=255). All GeoTIFFs are in South American Albers Equal Area projection at 10m resolution. For most use cases, 1 and 2 may be easier to analyze since they include reservoir location and surface areas without the added overhead of explicit reservoir shape information. For cases where the full reservoir shapes are needed, the GeoTIFFs can be analyzed via object-based methods such as skimage.ndimage.label.