The September 2020 White Volta flood: what the model and satellites show
I tested the flood maps behind my national assessment against the White Volta flood of September 2020, as mapped from Sentinel-1 satellite radar. In five 20 km squares the average agreement score was 0.58, where 1 is a perfect match. The population data put 6,371 people inside the observed flood. NADMO, Ghana’s national disaster agency, reported 7,561 people affected, but for the Upper East Region, a different area, so the two are context for each other, not a match.
- 3 districts holding the five test areas
- 257 other districts
How to read this: the outlined districts, Mamprugu Moagduri, Kumbungu and Savelugu, hold the centres of the five test areas along the White Volta north of Tamale. One flood, one region: a check on the method, not on every district.
- average agreement between the modelled and the satellite-observed flood, five test areas (1 = perfect)
- 0.58
- test areas of 20 × 20 km along the White Volta in northern Ghana
- 5
- people counted inside the satellite-observed flood; NADMO reported 7,561 affected in the Upper East Region, a different area
- 6,371
- of flooding seen by the satellite in September 2020
- 739.7km²
What happened on the White Volta in September 2020?
Heavy rain, together with water released from the Bagre dam in Burkina Faso from mid-August, flooded communities along the White Volta in northern Ghana. FloodList reported deaths in Bunkpurugu and in West and East Mamprusi. By 20 September NADMO had counted 7,561 people affected and 1,596 homes damaged in the Upper East Region.
Bagre is upstream, on the same river system. When its reservoir rises sharply, the operator releases water through the spillway, and that water flows on into Ghana. In 2020 FloodList described Mamprugu Moagduri, in North East Region, as badly affected.
For a flood modeller, the event had one big advantage: European Sentinel-1 satellites carry radar that sees through cloud, and the Copernicus Global Flood Monitoring service mapped the water at 20 m. That gave me an independent picture of where the river actually went.
How well did the model match the satellite flood?
Reasonably well. In five 20 km test areas the critical success index, the share of all flooded ground that both the model and the satellite called flooded, ranged from 0.43 to 0.71, with a mean of 0.58. A published test of global flood maps against observed floods reported means of 0.18 to 0.34.
I compared the satellite flood with the 1-in-10-year flood map from the JRC global river flood maps, the same maps I used for all 260 districts. The satellite map was 20 m; I counted a 90 m cell as flooded if at least a tenth of it was under water.
The edges are where the two disagree. In four test areas, nine in ten of the model’s misses were within about 370 m of flooding it did show. In test area 1, in Mamprugu Moagduri, the same measure was 2.8 km. No single correction would fix both.
In every test area, model and satellite agree on more than two-fifths of the flooded ground
How to read this: longer bars mean closer agreement. The dashed line is the best published average for a global flood model (JRC, 1-in-100-year maps); the shaded band is the range in a test of global models in Nigeria and Mozambique. Tap a bar to see how far the misses were from the flood edge.
Modelled 1-in-10-year extent (JRC global river flood maps v2) against the Sentinel-1 flood map (Copernicus Global Flood Monitoring), 20 m data counted on the 90 m grid.
Sources
My analysis (2025–26), also my Oxford MSc dissertation. Observed flood: Copernicus Emergency Management Service, Global Flood Monitoring (Sentinel-1). Benchmarks: Risling et al. (2024); Bernhofen et al. (2018).
| Test area (20 × 20 km) | Agreement (CSI) | Missed flooding: 9 in 10 within | Extra flooding: 9 in 10 within |
|---|---|---|---|
| Area 1 | 0.57 | 2.79 km | 0.51 km |
| Area 2 | 0.66 | 0.13 km | 0.45 km |
| Area 3 | 0.54 | 0.13 km | 0.37 km |
| Area 4 | 0.71 | 0.28 km | 0.33 km |
| Area 5 | 0.43 | 0.37 km | 0.37 km |
| Mean of the five | 0.58 | – | – |
Did the model count the right number of people and buildings?
Only as a rough check. The population data put 6,371 people inside the satellite flood footprint; NADMO reported 7,561 affected, but for the Upper East Region, a different area. For buildings, it depends on the flood edge: 474 buildings inside the footprint, but 3,073 if the edge moves out one 90 m cell, against 1,596 homes reported damaged.
The two people counts were not made for the same purpose or over the same ground. NADMO counted people affected in the Upper East Region; I counted people living under the water the satellite saw in my test areas, further downstream. So I read them only as a sign the population layer is in a plausible range, not as a match.
The buildings tell the more useful story. Moving the flood edge by a single cell, 90 m, multiplies the count six and a half times. Many homes along the White Volta sit close to the edge of the flood, so a small shift in where the water stops decides whether a whole settlement is counted.
| What was counted | From the model and satellite | Reported by NADMO, Upper East Region (via FloodList) |
|---|---|---|
| People | 6,371 in the satellite flood footprint | 7,561 affected |
| Buildings, flood edge as observed | 474 inside the footprint | 1,596 homes damaged |
| Buildings, edge moved out one 90 m cell | 3,073 inside the footprint | 1,596 homes damaged |
What did the test change in the national assessment?
It set the range of flood edges I tested everywhere. Because misses and over-reaches mostly fell within a few hundred metres, I ran every district with the edge pulled in two cells (180 m) and pushed out one (90 m). That moves Ghana’s modelled annual damage to homes from GHS 2.93 billion to 7.33 billion.
The same shift moves the number of people inside the 1-in-100-year flood from about 608,000 to about 1.78 million. Without the White Volta test, I would have had to guess that range.
The edge range then fed the test that matters for decisions: which districts stay in Ghana’s top 20 for modelled damage under all nine combinations of flood edge and damage curve. Thirteen do. That test, built on the range the White Volta flood gave me, is why I trust the short list more than any single run.
The short list: flood-prone districts in Ghana. How the flood edge moves the count of people: people at risk of flooding in Ghana.
What does the 2020 flood say about flood risk in northern Ghana?
That a ranking by money misses it. All three districts at the centre of the test areas are Monitor districts for modelled damage, yet Mamprugu Moagduri ranks 32nd in Ghana for people inside the 1-in-100-year flood. Homes there cost less to repair, so the floods count for less in cedis, not in lives.
This is why I recommend a people-based ranking beside the damage ranking. Northern districts along the White Volta would rise up a list that counts people, children and recovery capacity.
For early warning, flood maps and return periods are the risk knowledge a warning service starts from. Setting warning thresholds for the White Volta would need a gauged river model and agreed trigger levels. As a Flood Warning Duty Officer in England I worked with thresholds; I have not set them for Ghana.
The test areas sit in districts that rank higher for people than for damage
How to read this: the three outlined districts hold the test areas. Switch between people and damage and watch the north fade when the measure is money. Tap a district, or a row of the table, to read its figures.
Modelled river flooding, all 260 districts; people in the 1-in-100-year flood, and modelled damage to homes a year.
| District | Region | Test areas centred here | People in the 1-in-100-year flood extent | National rank for people | National rank for damage in that flood | Class |
|---|---|---|---|---|---|---|
| Mamprugu Moagduri | North East | 1 | 6,496 | 32nd | 94th | Monitor |
| Kumbungu | Northern | 2 | 3,075 | 55th | 97th | Monitor |
| Savelugu | Northern | 3, 4, 5 | 2,252 | 67th | 106th | Monitor |
Test-area districts found by placing each test area’s centre on the district boundaries; each 20 km square also reaches into its neighbours.
What can one flood not show?
One event, one regionA check on the method in northern Ghana. It does not prove that every district’s figure is right, and the lower Volta and the coast were not tested.
An unknown flood sizeNobody has established how rare the 2020 flood was. It need not match a 1-in-10-year flood.
What satellites missRadar images come every few days, so short-lived flooding between passes can be missed.
Dams not modelledThe Bagre releases are in the observed flood but not in the model’s flood maps, which do not represent how the dam is operated.
Not officialIndependent consultancy, also my Oxford MSc dissertation, presented for discussion with the World Bank. Not official figures of the Government of Ghana or NADMO.
Questions people ask
What is a critical success index?
A score for how well two flood maps agree: the area both call flooded, divided by the area either calls flooded. 1 is a perfect match and 0 is no overlap. A score of 0.58 means just over half of all the flooded ground was flooded in both.
Why compare a 1-in-10-year map with a real flood?
Nobody knows exactly how rare the 2020 flood was, so the test needs a modelled flood of broadly similar size. I used the 1-in-10-year map. The test checks where the model puts the water, not how often a flood that size comes.
What is the Bagre dam?
A dam in Burkina Faso, upstream of Ghana on the White Volta system. When its reservoir is high its operator releases water, and in 2020 those releases, from mid-August, added to the flooding in northern Ghana.
Does the test prove the district ranking is right?
No. It shows the flood maps put water in about the right places in one region, and it set the range of flood edges I tested. The ranking’s strength comes from holding under all nine combinations of assumptions. The ranking.
Can this check be repeated for other floods?
Yes. Copernicus now maps large floods from Sentinel-1 routinely, so the same test can be run for a later White Volta flood, the lower Volta in 2023 or the Tongu districts in 2026. More events would make the check much stronger.
Sources
- Nuhu, H. (2026). National fluvial flood-risk assessment and priority stability under uncertainty in Ghana. MSc dissertation, University of Oxford. Independent consultancy, presented for discussion with the World Bank. Method and results.
- Copernicus Emergency Management Service. Global Flood Monitoring, Sentinel-1 flood extent, September 2020. Global Flood Awareness System.
- European Commission Joint Research Centre. Global river flood hazard maps, version 2 (CC BY 4.0). JRC data catalogue.
- FloodList (8 September 2020). West Africa: floods hit Burkina Faso and northern Ghana. Article.
- FloodList (20 September 2020). Ghana: 5 dead after floods in Upper East Region (NADMO figures). Article.
- WorldPop, University of Southampton. Gridded population, 2026 (CC BY 4.0). worldpop.org.
- Google Research. Open Buildings v3 building footprints. Open Buildings.
- Risling, A., Lindersson, S. and Brandimarte, L. (2024). A comparison of global flood models using Sentinel-1 and a change detection approach. Natural Hazards 120, 11133–11152. doi:10.1007/s11069-024-06629-7.
- Bernhofen, M. V. et al. (2018). A first collective validation of global fluvial flood models for major floods in Nigeria and Mozambique. Environmental Research Letters. Record.
- Ghana Statistical Service / OCHA. Ghana subnational administrative boundaries (COD-AB). Humanitarian Data Exchange.
Work with me
Want to know whether a flood model matches what really happened?
I can test flood maps against satellite-mapped floods and reported impacts, and say plainly what the results can and cannot support. Remote or in-country.