IWMIHQ/africa_flood_occurrence

Africa Flood Occurrence

Annual flood occurrence across Africa at ~500 m, 2001 to 2024, from two parallel IWMI series on one grid: flood_occurrence (the finer count, per-year maximum 17 to 37) and inter_annual_flood_occurrence (consistently smaller). Both start at 1 and neither is documented by the publisher, but their distributions are those of counts. Rebuilt from IWMI Africa GeoPortal tile services, which serve no downloadable raster.

• Region Africa
• Period 2001-2024
• Theme Flood
• Grid EPSG:4326, 15942 x 17060 at 0.004491576 deg (~500 m)
• Bands 2
• Files 48 Cloud-Optimised GeoTIFFs under data/, keeping their source filenames, which encode the year and, where the product is tiled, the tile

Bands

Band	Type	Units	Nodata
flood_occurrence	int32	1	-2147483648
inter_annual_flood_occurrence	float32	1	-9999.0

Read it

These are COGs, so GDAL reads them in place over HTTP -- no full download, and a windowed read fetches only the bytes it needs:

from huggingface_hub import hf_hub_url
import rasterio

url = hf_hub_url("IWMIHQ/africa_flood_occurrence", "data/af_flood_2001.tif", repo_type="dataset")
with rasterio.open(url) as src:
    print(src.width, src.height, src.crs, src.nodata)
    patch = src.read(1, window=((0, 256), (0, 256)))

Or pull the whole repo:

from huggingface_hub import snapshot_download
path = snapshot_download("IWMIHQ/africa_flood_occurrence", repo_type="dataset")

Read this before using it

What the values MEAN is undocumented. There is no legend, no raster attribute table (the service answers "There is no raster attribute table available"), no description on the ArcGIS item, and nothing in the catalogue. The publisher's own tags -- "Flood, Occurrence, Inter-Annual" -- and three measured properties all point to a count of flood detections within the year rather than a class code: Values start at 1 and decay monotonically, with a stable shape in every one of the 24 years: 38-60% of valid cells are 1, 18-23% are 2, 8-14% are 3. The maximum is NOT fixed. It runs from 17 in 2001 to 37 in 2020, varying year by year -- a closed classification would have the same top code throughout, whereas a count naturally reaches higher in a worse year. (The service's own advertised maxValues of 17 is therefore per-service, not a scheme-wide limit; reading it once, off 2001, is how this product first got documented wrongly.) The wettest years by this measure are 2020, 2022 and 2024, which is consistent with the flooding those years are known for in eastern Africa. Still, this is inference from the data, not documentation. Confirm with IWMI before publishing anything that depends on it, and do not report "37 floods" as though the unit were established. The rasters are extremely sparse and that is expected, not a defect. Only 0.007% to 0.054% of the 271,970,520 cells carry a value -- 17,877 cells in 2002, 147,486 in 2020 -- because only cells where flooding was detected are populated and everything else is nodata. So every statistic must be computed over the valid cells alone; a mean over the frame is meaningless, and a "percentage of Africa flooded" taken from cell counts without honouring the nodata mask will be wrong by orders of magnitude. That eightfold swing in populated cells between years is the product's main signal, and it is also a warning: because the count of valid cells varies so much, any comparison between years must state whether it is comparing flooded AREA (valid cell count) or flood INTENSITY (the values themselves). They move differently -- 2003 and 2019 have similar areas but different maxima. Nodata is the int32 minimum, -2147483648, not 0. The services report a minimum value of 1, so 0 is outside the real range -- but it is also the value an unmasked read would plausibly produce, and treating "no flood detected" and "not assessed" as the same thing is exactly the error this avoids. Permanent water appears to be excluded. Sampled hits concentrate where large floods are expected -- the Niger inland delta, the Lower Shire, the Sudd, Lake Chad, the Barotse floodplain -- but the Okavango delta returns none at all, which is consistent with permanent wetland being masked out of a flood product. That means absence of a value does not mean absence of water. THE TWO BANDS ARE NOT INTERCHANGEABLE, and their relationship was measured across all 24 years and 1,462,852 shared cells rather than sampled. Both are integral counts -- inter_annual_flood_occurrence is stored float32 but every value is a whole number from 1 to 11, so do not read the float type as precision. Comparing them cell by cell: 54.8% of shared cells hold exactly the same value. flood_occurrence is larger in about 43%, by +1 in 17.4% of cells and decaying from there. And in 42,161 cells -- 2.9% -- flood_occurrence is SMALLER, always by exactly one, never by more. So the safe statement is flood_occurrence >= inter_annual_flood_occurrence - 1; neither band dominates the other outright. A first check on two tiles suggested flood_occurrence was never smaller at all. That was wrong, and it is recorded here because the difference matters: a monotonic relationship would imply one band is a strict refinement of the other, and it is not. Their footprints are identical in 21 of the 24 years and differ in 2002, 2005 and 2022 by 1,359 to 1,618 cells. So the valid mask is per-band, not per-dataset. One unexplained structural feature: the maximum ratio between the bands is exactly 4.0 in 23 of the 24 years, and 7.0 in 2022. A cap that lands on a round number in 23 years running is unlikely to be coincidence, but nothing published explains it, so it is noted rather than interpreted. Rebuilt from a tile cache rather than downloaded, so the values are whatever was published into the LERC pyramid. LERC is lossless and the level used is the service's native resolution, but the ArcGIS service recorded in each file's source tag stays the authority. "_Clean" in the source service names implies a filtering or QA step that is not described anywhere. Whatever was removed, and on what basis, is unknown.

Provenance

Rebuilt from IWMI Africa GeoPortal ArcGIS image services, which serve no downloadable raster: they answer exportImage with HTTP 400 because their capabilities are Image,TilesOnly. The full-resolution grid was recovered by fetching every LERC tile at the pyramid level matching the service's own native resolution, decoding, and mosaicking. Each GeoTIFF carries the originating service in its source tag, and that service -- not this copy -- is the authority for its values.

The same product is indexed in IWMI's Open Data Cube, which is the route to its footprint, time extent and band metadata as STAC:

https://explorer-production-0070.up.railway.app/stac/collections/africa_flood_occurrence

Citation

International Water Management Institute (IWMI), Africa Flood Occurrence. Published via the IWMI Open Data Cube. Licence CC-BY-4.0.

@misc{iwmi_africa_flood_occurrence,
  title        = {Africa Flood Occurrence},
  author       = {International Water Management Institute},
  howpublished = {Hugging Face Hub, IWMIHQ/africa_flood_occurrence},
  note         = {Derived from IWMI Africa GeoPortal services; see Provenance},
  year         = {2026}
}