Session: 2026-06-18

← Index

Summary

Comparative analysis of the 2025 and 2026 Adamawa action triggers, and exploration of GloFAS reforecast configuration options for the 2026 readiness trigger. Evaluation uses Floodscan SFED annual maxima (1998–2023) as ground truth and the Weibull plotting position formula for empirical return periods throughout.

2025 vs 2026 trigger comparison

Trigger definitions

The two triggers differ in exactly one year each: 2025 fires in 2003 (a 3-yr Floodscan event, not a 4-yr event); 2026 fires in 2018 (a 4-yr Floodscan event). Both miss 2015 and 2023.

Performance metrics (reanalysis, 1998–2023, Weibull RP)

Evaluated against Floodscan event years: 3-yr = {1998, 1999, 2003, 2012, 2015, 2018, 2019, 2022, 2023}; 4-yr = {1999, 2012, 2015, 2018, 2022, 2023}; 5-yr = {1999, 2012, 2015, 2022, 2023}.

FPR = FP / (FP + TN). At 3-yr and 5-yr benchmarks the triggers are statistically identical. The 2026 trigger is clearly better at the 4-yr level — its designed threshold — because it fires in 2018 (a genuine 4-yr event) instead of 2003 (only a 3-yr event).

Qualitative comparison

2025 strengths

• Longer potential lead time via GloFAS reforecast component (can fire days to weeks ahead of peak)
• Two independent model signals (GloFAS + GRRR) — a failure or bias in one doesn’t cause a miss
• Simpler to monitor operationally (single station)

2025 weaknesses

• Single point of failure: everything depends on Wuroboki; gauge malfunction or events peaking elsewhere cause misses or false alarms
• OR logic means false alarms from either source are additive
• GloFAS threshold (~5-yr RP) is a high bar; in practice the trigger likely fires mostly through the GRRR component, making GloFAS’s contribution smaller than intended

2026 strengths

• Spatially robust: requiring 6 of 10 gauges to agree is a strong quality filter; local noise at a single station doesn’t trigger activation
• Internally consistent design: all gauges use the same RP threshold from the same data source
• Better calibrated to its 4-yr RP target, with better metrics at that level

2026 weaknesses

• Short reforecast evaluation period (2016–2022 only, 7 years) — insufficient to draw strong conclusions about reforecast lead times
• No long-range forecast component: maximum early warning bounded by GRRR reforecast horizon (≤7 days), tight for pre-positioning
• Misses 2003 (moderate flood year caught by 2025) and both miss 2015 and 2023

Selected GRRR gauges — Adamawa

Ten gauges selected by Spearman ρ rank (daily correlation with Floodscan SFED, Aug–Nov, lag range −7 to +14 days). All are GRRR (Google) source. Trigger fires if ≥6 of 10 gauges simultaneously exceed their individual 4-yr empirical RP threshold on at least one wet-season day.

QV = quality-verified by Google. 4-yr threshold = empirical Weibull RP from wet-season annual maxima, 1998–2023. n trigger yrs = number of the 6 reanalysis trigger years in which this gauge contributed (i.e. was one of the ≥6 simultaneously exceeding on the trigger day).

Notes on hybas_1120840690

This gauge is a clear outlier: its 4-yr threshold is 142.7 m³/s, roughly 8× lower than the other nine gauges (~1,100 m³/s). It also has a positive lag (+2 days), meaning it peaks after the Floodscan SFED peak rather than before, and contributed to only 4 of 6 trigger years. It likely represents a smaller tributary rather than the main Benue channel. Despite the low threshold, it is not disqualifying in the ≥6-gauge logic since it only provides one vote out of ten.

Google official thresholds (for reference)

The trigger uses empirical Weibull RP thresholds, not Google’s official levels. The Google model provides warning / danger / extreme danger thresholds (approximately 2-, 5-, and 20-yr RP) for gauges with quality-verified models:

The empirical 4-yr thresholds (used in the trigger) are very close to the Google warning levels for most gauges (~1,078–1,117 vs ~1,078–1,095 m³/s), consistent with warning level ≈ 2-yr RP and danger level ≈ 5-yr RP in Google’s framework.

Empirical return periods

Computed using the Weibull formula: RP = (n + 1) / k.

The readiness trigger should have a shorter RP than the action trigger to reflect its role as a more sensitive, lower-cost signal. Options A and B both satisfy this (3.0 yr and 3.5 yr vs action’s 4.5 yr).

GloFAS readiness trigger configuration options

The readiness trigger uses the GloFAS reforecast ensemble mean at Wuroboki. A year fires if the ensemble mean exceeds the threshold at any leadtime ≤ MAX_LT in the wet season (Aug–Nov). Evaluated over 2003–2022 (n = 20 years). Action trigger years with reforecast coverage: 2012, 2018, 2019, 2022 (4 years) — all included in POD.

Option A — 3,132 m³/s, LT ≤ 13d

Lead times vs action trigger: 2012 +2d, 2019 +46d, 2022 +13d, 2018 not detected.

Option B — 3,250 m³/s, LT ≤ 12d

Lead times vs action trigger: 2012 −15d (fires after), 2019 +46d, 2022 +3d, 2018 not detected.

Threshold summary

The key trade-off is whether to capture 2012. The maximum GloFAS ensemble mean (LT≤13d) before the Aug 23 action date in 2012 was only 3,142 m³/s at LT=1d — meaning any threshold above 3,142 m³/s fires after the action trigger in 2012. The +2d lead in option A is therefore operationally marginal (nowcast, not forecast). Both options fail to detect 2018; this is an inherent GloFAS forecasting limitation rather than a threshold choice. Both options provide strong early warning in 2019 (+46d) and adequate warning in 2022 (+3–13d depending on option).

Operational considerations and limitations for stakeholders

GloFAS reforecast lead time is uneven across years

Lead times vary substantially: 46 days in 2019, 2–13 days in 2022, and essentially zero in 2012. Stakeholders should not expect a consistent preparation window — in some years the readiness trigger fires weeks early, in others it fires at the same time as the action trigger or not at all. The 2019 result should not be used as the typical expectation.

GloFAS provided no meaningful early warning in 2012

In 2012, the GRRR multi-gauge action trigger (Aug 23) fired before or at the same time as any useful GloFAS signal. The flood came on quickly and was not forecast far in advance by GloFAS. This is a known limitation of single-basin ensemble forecasts for flash-type events on the upper Benue.

False alarm rate at 5% cost is manageable but should be communicated

At the recommended threshold, the readiness trigger fires in approximately 1 of every 3–3.5 years, compared to the action trigger’s 1 in 4.5 years. Roughly 3–4 readiness activations per 20 years will not be followed by an action trigger activation. At 5% funding, this is financially acceptable, but stakeholders should be briefed that a readiness activation does not guarantee a full activation will follow.

2018 is not detected by GloFAS at any practical threshold within LT≤13d

2018 is a 4-yr Floodscan event and an action trigger year, but the GloFAS reforecast ensemble mean only exceeds the threshold at LT≥14d in that year. GloFAS appears to have underforecast the 2018 event at shorter leadtimes. The readiness trigger cannot reliably serve as an early warning for 2018-type events.

Both 2015 and 2023 are missed by both action and readiness triggers

These are genuine 4–5-yr flood years that neither trigger configuration detects. Investigating why these years are not captured (different flood pathway, upstream conditions, gauge coverage) is a priority for improving future trigger design.

Short reforecast evaluation period limits confidence

The GloFAS reforecast only covers 2003–2022 (20 years), and the GRRR multi-gauge reforecast (2016–2022, 7 years) is too short to draw robust conclusions about reforecast lead times at the action trigger level. All reforecast performance figures carry substantial uncertainty.

The 2025 FAR figures in documentation (18%, 10%) differ from values computed here

The 2025 documentation reported FAR of 18% (3-yr) and 10% (5-yr). These likely reflect a longer evaluation period extending before 1998 (where pre-Floodscan trigger activations were counted as false alarms), a different event year set derived from a different Floodscan pixel dataset, and/or rank-based rather than Weibull RP thresholds. The figures are not directly comparable to the metrics computed here.

Reference: key discharge thresholds at Wuroboki (GloFAS reanalysis, empirical Weibull, 1998–2022)