FloodWatch Ghana

← Back to map

Risk Methodology & Model Validation (v0.1)

Greater Accra Region, Ghana  ·  Updated April 2026

This document provides a comprehensive overview of the FloodWatch Ghana risk model, the district risk leaderboard, engineering history, and a full quantitative and qualitative validation of both model versions against the May 18, 2025 Greater Accra flood event.


1. The Risk Model (v0.1)

FloodWatch Ghana v0.1 is a structural baseline model. It identifies areas chronically prone to flooding based on their physical and environmental characteristics — terrain, drainage, land cover, and rainfall patterns.

1.1 Weighted Composite Formula

Each 30m pixel is assigned a risk score from 0 (Low) to 1 (High) using a weighted combination of five input layers:

ComponentWeightDirectionSourceRationale
Elevation30%InvertedNASA SRTM (30m)Low-lying areas are natural catchments for surface runoff.
Precipitation25%NormalGPM IMERG Final Run (0.1°)Actual observed monthly rainfall drives runoff volume.
Terrain Slope20%InvertedDerived from SRTMFlat terrain cannot drain quickly and pools surface water.
Imperviousness15%NormalESA WorldCover (10m)Paved and urban surfaces prevent infiltration into soil.
Water Proximity10%InvertedOpenStreetMapProximity to rivers and drainage channels increases inundation risk.

Formula:

Risk = 0.30×(1−DEM_norm) + 0.25×Rain_norm + 0.20×(1−Slope_norm) + 0.15×Imperv_norm + 0.10×(1−Water_norm)

All input layers are min-max normalised to [0, 1] before compositing. The final composite is reclassified using percentile-based stretching (p25 and p75 breakpoints) to distribute risk scores across the full [0, 1] range and avoid compression in the middle.

1.2 Rainfall Data Source — Why GPM IMERG over ERA5/CHIRPS

The rainfall layer is the most operationally significant input to update. Two approaches have been used across model versions:

SourceTypeLatencyAccuracyUsed in
CHIRPS v2.0Climatological meanDaysModeratev0.1 original
ERA5-LandReanalysis meanDaysGoodv0.1 original fallback
GPM IMERG Final RunActual monthly observed~3.5 monthsBest (gauge-corrected)v0.1 recalculated
GPM IMERG Late RunNear real-time~12 hoursGoodv0.1 recalculated fallback

CHIRPS and ERA5 return the same climatological average for June regardless of the year — June 2019 and June 2024 produce identical values. GPM IMERG returns the actual measured precipitation for that specific month, making the model genuinely responsive to real rainfall conditions. The v0.1 recalculated model uses GPM IMERG Final Run for June 2024 (198 mm/month mean over Greater Accra).


2. District Risk Leaderboard

2.1 Original Model (CHIRPS/ERA5 rainfall, no percentile reclassification)

Mean risk scores computed per district from the original pipeline run.

RankDistrictMean RiskMax RiskFlooded May 2025
1Ablekuma West0.83980.9928No
2Weija Gbawe0.81230.9957Yes
3Ga Central0.72580.9796No
4Accra Metropolis0.71700.8681Yes
5Ga West0.70630.9091No
6Ga South0.70350.9168No
7Ablekuma North0.69220.9129No
8Ablekuma Central0.68760.9673No
9Ayawaso East0.67410.8010No
10Korle-Klottey0.67080.8220No
11La-Dade-Kotopon0.66650.8261No
12Ayawaso North0.64280.7923No
13Okaikwei North0.62160.7901No
14Ayawaso Central0.60240.7815No
15Krowor0.59940.7714No
16Ledzokuku0.56330.7841No
17Ayawaso West0.55550.7763No
18Ga East0.55290.8011Yes
19Ga North0.53990.8156No
20Tema0.52150.7691Yes
21Tema West0.48870.7598Yes
22Ningo-Prampram0.46461.0000No
23Ada East0.46350.8083No
24La-Nkwantanang-Madina0.46190.7043Yes
25Adenta0.46100.7331Yes
26Kpone-Katamanso0.45070.7605No
27Ada West0.43760.7787No
28Shai Osudoku0.42230.9061No
29Ashaiman0.36540.6527No

2.2 Recalculated Model (GPM IMERG rainfall + percentile reclassification, April 2026)

RankDistrictMean RiskFlooded May 2025
1Ayawaso North0.9301No
2Ledzokuku0.9204No
3Krowor0.9181No
4Tema West0.8956Yes
5Ashaiman0.8941No
6La-Dade-Kotopon0.8910No
7Ga Central0.8814No
8Ablekuma North0.8734No
9Ayawaso West0.8686No
10Ayawaso East0.8682No
11Okaikwei North0.8570No
12Ayawaso Central0.8525No
13Korle-Klottey0.8525No
14Adenta0.8408Yes
15Tema0.8357Yes
16Ga West0.8160No
17Ablekuma Central0.8157No
18Ga East0.8114Yes
19Accra Metropolis0.7847Yes
20Weija Gbawe0.7570Yes
21La-Nkwantanang-Madina0.7531Yes
22Ga South0.7523No
23Ga North0.7507No
24Kpone-Katamanso0.7430No
25Ablekuma West0.7418No
26Ningo-Prampram0.4870No
27Ada East0.4747No
28Shai Osudoku0.4541No
29Ada West0.3971No

3. Validation — May 18, 2025 Flood Event

3.1 Event Summary

On May 18, 2025, Greater Accra experienced a severe flash flooding event following approximately 132mm of rainfall in a short period — roughly the equivalent of a full month's rain in a single day. The event caused widespread flooding across multiple districts. Reported flooded districts (sourced from The Watchers, GDACS, and Copernicus EMS):

Flooded (7 of 29 districts): Weija Gbawe · Accra Metropolis · Ga East · Tema · Tema West · La-Nkwantanang-Madina · Adenta

Not flooded (22 districts): All remaining districts.

3.2 Quantitative Metrics — Model Comparison

Mean Risk Score by Flood Status

MetricOriginal ModelRecalculated ModelVerdict
Mean risk — flooded districts0.57360.8112Recalc higher ✓
Mean risk — non-flooded districts0.59530.7745
Difference (flooded − non-flooded)−0.0217+0.0367Recalc correct direction ✓
% flooded districts flagged High Risk (≥0.70)28.6% (2/7)100% (7/7)Recalc better ✓
% non-flooded districts flagged High Risk (≥0.70)18.2% (4/22)81.8% (18/22)Original more precise

Confusion Matrix at 0.70 Threshold

Original ModelRecalculated Model
True Positives (flooded, flagged high)27
False Positives (not flooded, flagged high)418
True Negatives (not flooded, flagged low)184
False Negatives (flooded, missed)50
Precision0.330.28
Recall0.291.00
F1 Score0.310.44

3.3 Qualitative Assessment

Original Model

The original model correctly placed two of the most historically flood-prone districts — Weija Gbawe (rank 2) and Accra Metropolis (rank 4) — in its top tier. These are well-known chronic flood zones in Greater Accra and their high ranking reflects genuine structural risk (low elevation, dense impervious surfaces, proximity to the Odaw River and Korle Lagoon drainage system).

However, the model missed five flooded districts entirely at the 0.70 threshold:

This is the model's most significant qualitative failure. Adenta and La-Nkwantanang-Madina are peri-urban and inland districts that were overwhelmed by the volume of the May 2025 event — their structural characteristics (moderate slope, mixed land cover) do not mark them as chronic flood zones, but a 132mm single-day rainfall event overloaded their drainage regardless. The original model, built on climatological rainfall averages, had no mechanism to capture this.

The mean risk of flooded districts (0.574) was actually lower than non-flooded districts (0.595) — the model ranked flooded areas as marginally safer on average. This is a fundamental failure of direction.

Recalculated Model

The recalculated model shows a meaningful improvement. With GPM IMERG actual rainfall (June 2024, 198 mm/month mean) and percentile reclassification applied:

The main weakness of the recalculated model is low precision (0.28): 18 of 22 non-flooded districts also score above 0.70. The score distribution is compressed into a narrow high band (most districts fall between 0.74–0.93), making it difficult to discriminate flooded from non-flooded at the district mean level alone. The bottom four districts — Ningo-Prampram, Ada East, Shai Osudoku, Ada West — are correctly identified as low risk; these are predominantly rural and coastal areas with very different terrain and land cover.

3.4 Overall Verdict — Which Model Performs Better?

The recalculated model is the stronger performer.

CriterionOriginalRecalculatedWinner
Direction of risk signalWrong (flooded < non-flooded)Correct (flooded > non-flooded)Recalculated
Recall — flooded districts caught0.291.00Recalculated
F1 Score0.310.44Recalculated
Precision0.330.28Original (marginally)
Qualitative alignment (known flood zones)Partial (2/7)Strong (7/7)Recalculated
Score discrimination across districtsBetter spreadCompressed mid-highOriginal
Rainfall data qualityClimatological averageActual observedRecalculated

The recalculated model wins on every meaningful criterion except precision. Its near-zero false negative rate is critical for a flood risk application — missing a flooded district is a worse failure than over-flagging a safe one. The original model's apparent precision advantage is misleading: it achieved it by simply scoring most districts as moderate risk, meaning it also missed five of the seven districts that actually flooded.

The precision gap (0.28 vs 0.33) is a known structural limitation of both models. A static weighted composite applied at the district mean level will always have difficulty separating flash-flood-driven events from structural risk — the underlying issue is that the May 2025 event was an extreme single-day episode, while the model represents chronic susceptibility. Improving precision requires dynamic, event-driven inputs.


4. Engineering History & Bug Resolutions

4.1 The "Global Average" Bug (0.508)

During early development, every district incorrectly displayed a uniform Mean Risk Score of 0.508.

4.2 Rainfall Source Upgrade (CHIRPS → GPM IMERG)

The original model ingested rainfall from CHIRPS v2.0 or ERA5-Land — both climatological products that return the same historical average regardless of the actual year processed. This meant the model could not respond to unusually wet or dry months.

The pipeline was upgraded to use NASA GPM IMERG as the primary source, with a 4-tier fallback chain:

GPM IMERG Final Run  →  GPM IMERG Late Run  →  ERA5-Land  →  CHIRPS v2.0

GPM IMERG Final Run is bias-corrected against ground rain gauges and available with approximately 3.5 months latency. For June 2024, the actual observed mean rainfall over Greater Accra was 198 mm/month (range: 126–300 mm/month across the region), compared to the climatological average which does not vary by event.

4.3 Percentile Reclassification

The original pipeline applied min-max normalisation directly to the composite score, which resulted in compressed mid-range scores across most districts. The recalculated model adds a percentile-based reclassification step using the 25th and 75th percentile breakpoints of the pixel-level risk distribution:

score < p25          →  mapped to [0.00, 0.33]   (low tier)
p25 ≤ score < p75   →  mapped to [0.33, 0.67]   (moderate tier)
score ≥ p75          →  mapped to [0.67, 1.00]   (high tier)

This better utilises the full output range and sharpens the separation between low, moderate, and high risk areas at the pixel level — though district mean compression remains at the 0.70+ band for most urban districts.

4.4 COG Pipeline & Tile Serving

All risk outputs are served as Cloud-Optimised GeoTIFFs (COG) from Google Cloud Storage, rendered via TiTiler. Previous versions encountered issues with:


5. Limitations & Roadmap

5.1 Current Limitations

5.2 Future Roadmap (v1.1)

FeatureDescriptionImpact
Dynamic risk layer Real-time GPM IMERG rainfall thresholds triggering risk score adjustments on the day of an event High — addresses the core precision gap
Sentinel-1 SAR validation Flood extent mapping via Google Earth Engine for quantitative spatial accuracy metrics beyond district means High — enables pixel-level validation
Drainage infrastructure layer OSM and NADMO drainage network data as an additional composite input Medium
Property-level API Dynamic backend for individual parcel risk queries Medium
Temporal validation Rerun model with May 2025 GPM data to validate under event-matched rainfall Medium