Somalia floods: data-source review — Somalia Riverine Flood Trigger

1-in-3.2 envelope activation rate (target 1-in-3)

7 of 8 severe years caught (1-in-5 or rarer at the gauge)

4 + 3 gauges monitored: Juba and Shabelle

7 days lead time | sources: GloFAS v5 + Google GRRR

1Key takeaways

Model performance differs by river and season. Google carries the two Gu windows and GloFAS v5 the two Deyr windows. GloFAS v5 alone cannot reach the 1-in-3 activation rate (its nearest configuration fires once every 2.9 years), which is what puts Google in the mechanism.
Accuracy holds from one to seven days ahead, so a one-week action window is defensible. Models still miss a share of floods, so SWALIM guidance belongs in the trigger design.
The gauge record is the reference, and only the reporting gauges can serve a live trigger: 4 on the Juba, 3 on the Shabelle. Thresholds are fitted on post-2000 data, because flooding has been recorded far more often in recent decades.
A flood is defined by the computed 1-in-3-year level at the gauge, not by an official SWALIM mark: it is frequency-calibrated, so it means the same thing at every gauge.

2SWALIM thresholds

SWALIM publishes three risk levels per station: moderate, high and bank full. Each station's 1-in-3-year level is estimated from its own annual maxima and compared against them. Where the 1-in-3-year level sits above the moderate mark, that mark is crossed more often than once every three years.

Where the 1-in-3-year level sits against the official marks. The official Moderate band means very different things from gauge to gauge, which is why the trigger monitors the computed RP3 instead.

3SWALIM events

An event is a spell in which the level sits at or above the baseline. The level must stay below it for at least 14 days before a new crossing counts separately, so a brief dip mid-flood does not split one flood in two. Readings stop rising once a gauge reaches bank full, so the largest events are understated.

What the two-gauge benchmark sees, and what it misses. A river-season counts as a flood only when two or more of that river's gauges cross their own level, which keeps one record from deciding the benchmark but has three consequences worth stating.

It agrees with the impact record on the big years. All five costliest years in EM-DAT and CERF terms — 2006, 2018, 2019, 2020 and 2023 — are severe under the rule.
1999 to 2001 cannot be assessed. The Juba had no gauge reporting and the Shabelle only one, so no consensus is possible and those years read as quiet rather than as unknown. Nothing activates before 2005, so no year is wrongly scored as a miss, but the benchmark effectively starts in 2002.
2021 is a genuine gap. EM-DAT records 400,000 people affected, yet only Bardheere on the Juba and Belet Weyne on the Shabelle crossed, so the rule reads no flood. Belet Weyne's peak that year sits exactly at bank full (8.30 m), where the gauge record is censored and the true level is unknown, so consensus can be understated in exactly the years that matter most.
2016 is the reverse case. Three of the four Juba gauges and all three Shabelle gauges crossed their levels in Gu, but EM-DAT records nobody affected. Gauge levels and recorded impact are not the same measurement.

Seven gauges still reporting: four on the Juba, three on the Shabelle.

The two baselines side by side. Amber cells are years the official Moderate mark activated and the RP3 did not, which is where the two definitions disagree.

4SWALIM risk level crossings

How often each gauge crosses its RP3 level. Most gauges reach it in roughly one season-year in three, by construction.

5SWALIM against flood exposure

The gauge record matters because of who sits behind it: the districts along both rivers with the largest populations in flooded areas.

6Reanalysis performance

Each model's 1-in-3-year signal against the events recorded at the gauges, 2002 to 2023, within a 7-day window. POD is the share of events caught; FAR the share of alarms that were false.

Each model's 1-in-3-year signal against the floods the gauges actually recorded, within a 7-day window.

Station	Events	GloFAS v5	GloFAS v4	Google GRRR	GEOGloWS v2
dollow	3	1.00 / 0.40	0.67 / 0.60	1.00 / 0.50	0.67 / 0.60
luuq	9	0.44 / 0.50	0.22 / 0.71	0.44 / 0.43	0.11 / 0.88
bardheere	4	0.50 / 0.75	0.25 / 0.86	0.50 / 0.75	0.25 / 0.88
bualle	7	0.29 / 0.67	0.43 / 0.57	0.43 / 0.50	0.43 / 0.57
belet_weyne	10	0.60 / 0.14	0.10 / 0.88	0.50 / 0.29	0.30 / 0.62
bulo_burti	9	0.44 / 0.43	0.11 / 0.88	0.33 / 0.57	0.44 / 0.50
jowhar	10	0.20 / 0.71	0.10 / 0.88	0.10 / 0.86	0.10 / 0.86

7Model correlation

Daily tracking rather than event detection: the best-lag rank correlation between each model and the river's reference gauge, allowing for travel time.

River	Best	Second	Third
Juba (worse season)	GloFAS v5 0.83	GloFAS v4 0.82	GEOGloWS v2 0.8	Google GRRR 0.53
Shabelle (worse season)	GloFAS v4 0.78	GloFAS v5 0.74	GEOGloWS v2 0.73	Google GRRR 0.63

8Forecast correlation

The same test on the forecasts rather than the hindsight simulations, leads 1 to 7.

The same test run on the forecasts rather than the hindsight simulations. Accuracy holds across leads 1 to 7, which is what makes a one-week action window defensible.

9Selected stations

River	Gauges used
Juba	dollow, luuq, bardheere, bualle
Shabelle	belet_weyne, bulo_burti, jowhar

Gauges whose records stop in 2008 or earlier can calibrate but cannot monitor. Upstream stations see the flood before the downstream reference gauge: Dollow leads Luuq by about five days in Deyr, and Belet Weyne leads by about six in Gu.

10Grid search

Every combination of station return period and number of gauges that must agree, scored against the gauge benchmark. The outlined cell is the one adopted.

Every combination of station return period and number of gauges that must agree, scored against the gauge benchmark. The outlined cell is what we adopted.

11Best configuration

River	Season	Forecast	Gauge threshold	Gauges that must agree	Would have activated in
Juba	Gu	Google GRRR	1-in-5	3 of 4	2013, 2016, 2018, 2020
Juba	Deyr	GloFAS v5	1-in-4	3 of 4	2006, 2023
Shabelle	Gu	Google GRRR	1-in-6	2 of 3	2013, 2016, 2018, 2020
Shabelle	Deyr	GloFAS v5	1-in-4	2 of 3	2006, 2013, 2014, 2019, 2020, 2023

The full amount is released whenever the trigger is reached along either river in either season, so the combined rate across all four windows is what the budget must be sized on. As configured it would have released in 8 of 25 years, once every 3.2 years, catching 7 of the 8 severe years.

GEOGloWS is excluded from the adopted design. Its forecasts run below its own retrospective, so thresholds would have to be refitted on its forecast archive, which only begins in July 2024 and can neither be backtested at lead time nor validate a debias. Its per-gauge detection was also the weakest of the four models, and the SFDC bias correction lowers it further rather than fixing it.

Year by year, what the adopted configuration would have done, and what the envelope does when any of the four windows activates.