New Google model: re-selecting the gauges

1What changes

• The gauges. None of the ten endorsed gauges make the new top ten. The new ten sit further down the Benue: 7 downstream of Numan, where the Gongola joins, and 3 between Numan and Yola. All ten endorsed gauges sit upstream of them, from about 15 km below Yola to above it, and rank 12 to 29 on the new model.
• The ranking is almost flat. On the new model the top 15 candidates all correlate with FloodScan between 0.757 and 0.765, and the tenth and eleventh differ by only 0.0004, so which ten are picked comes down to the fourth decimal. On the old model the top ten ranged from 0.711 to 0.742. The new model also matches FloodScan best with no time offset at every one of these fifteen gauges.
• Activation years. 1998, 1999, 2012, 2018, 2020, 2022 with the re-selected gauges, against 1998, 1999, 2012, 2018, 2019, 2022 with the endorsed gauges (the same on either model). 2019, a FloodScan 1-in-3 year, drops out; 2020, below the FloodScan 1-in-3 level, comes in. Against the design's 1-in-4 benchmark the scores are unchanged; against 1-in-3 there is one hit fewer and one false alarm more.
• Timing. First day the rule is met, re-selected gauges against the old design: same day in 1998, 7 days earlier in 1999, 23 days later in 2012, 1 day earlier in 2018, 5 days earlier in 2022.
• Forecasts (complete seasons 2016–2022): the re-selected gauges meet the rule in 2018, 2020, 2022, against 2018, 2019, 2022 on the old design and 2016, 2018, 2019, 2022 on the new model with the endorsed gauges.
• Levels. The re-selected gauges carry more water: 1-in-4 levels of 3,702–4,341 m³/s, against 3,462–3,556 for the endorsed main-channel gauges on the new model. The small tributary gauge of the endorsed set drops out.

2Which gauges

Outlets of the 42 Google candidates (hover for ranks). Filled: the ten re-selected on the new model. Open rings: the ten endorsed gauges.

3When the trigger is met

Rows as on the comparison page: FloodScan class, then for each set-up the most gauges over their own 1-in-4 level on any one day (filled where 6 or more), then forecast years (complete seasons 2016–2022, leads 0–7). Hover a cell for details.

Daily flow at the sixth-highest of each set-up's ten gauges as a share of that gauge's own 1-in-4 level, August to November; the rule is met above 100%, and dots mark the first such day. FloodScan: flooded fraction in Adamawa as a share of its own 1-in-4 level. Years shown: every year from 1998 to 2023 in which any set-up's record or forecast met the rule.

4Performance against FloodScan

Scored as in the design. Record: 1998–2023. Forecast: complete seasons 2016–2022.

5Levels of the re-selected gauges

1-in-4 level on the new model's August–November maxima 1998–2023 (m³/s).

6Method

• Selection as in the design's workflow notebooks 01, 03 and 04: the stored candidate pool; each Google candidate's best Spearman correlation with FloodScan (August–November 1998–2023, lags of −3 to +14 days); all candidates ranked together, top ten kept. GloFAS at Wuroboki keeps its stored correlation (0.679), as its series does not depend on the Google model.
• Run on the old model, the same code reproduces the stored correlations of all 42 Google candidates exactly and selects the ten endorsed gauges.
• Trigger and scores with the comparison page's code (analysis.py); this variant in reselect.py.