As part of the GloFAS v2.1 30-day upgrade, the way the river discharge thresholds were calculated was reviewed. In this document the changes in the threshold computation and the datasets used is described. In addition, the differences between GloFAS v2.1 and v2.0 thresholds are highlighted for the 5-year return period as an example.

GloFAS flood threshold computation

In GloFAS 30-day, thresholds are represented by return period discharge magnitudes. These thresholds are computed in the following way:

• Thresholds are computed from a discharge reanalysis over a long period of several decades. Currently the ERA5 climate reanalysis of ECMWF is used to produce the GloFAS discharge reanalysis.
  
• Discharge annual maxima are extracted from the daily time series of the discharge reanalysis. E.g. for a 40-year period this will result in 40 annual maxima values.
  
• A Gumbel distribution is fitted to these annual maxima time series using the L-moment method.
• The river discharge magnitudes are identified that correspond to the 2-, 5- and 20-year return periods. These are the GloFAS flood thresholds.

Difference in estimated flood return levels

GloFAS v2.1 and GloFAS v2.0 flood thresholds differ because of three aspects:

• Updated reanalysis: GloFAS v2.1 is benefiting from the officially released ERA5 reanalysis, while GloFAS v2.0 could only use the pre-released ERA5, which included some non-quality-assured data periods.
  
• Longer period: GloFAS v2.1 has a slightly extended period of 1979-2018 to use, while GloFAS v2.0 was based on the period of 1981-2017.
• Minor methodology revision: GloFAS v2.1 uses all available years (40 years in total) in the extreme value fitting. In contrast, GloFAS v2.0 removed years which had annual maxima below the 37-year discharge mean threshold, and fitted a distribution with a minimum of 5 years of data.
  

The impact of the new estimated return levels was examined and is presented here for the 5-year return magnitudes, Q5v2.1 and Q5v2.0 (Figure 1). Results are shown as proportional difference against Q5v2.0 value as defined by (Q5v2.1 - Q5v2.0) / Q5v2.0. Positive (negative) values showing a larger (smaller) v2.1 threshold compared with v2.0, with a value of 0 showing identical results,  a value of 1 showing Q5v2.1 equal to twice Q5v2.0 and a value of -0.5 showing Q5v2.1 equal to half Q5v2.0. Three patterns could be identified:

• Large part of the river network above 1000 km² shows limited differences (less than 10% difference, light grey rivers).

• In dry areas (e.g. Sub-Saharan Africa, large parts of Australia, Saudi Arabian Peninsula, Pakistan, and the central parts of the US), Q5v2.1 threshold is substantially lower than Q5v2.0. This is because Q5v2.0 was calculated from a sample with many dry years removed before fitting of the extreme value curves, hence not representing the 5-year return period likelihood but a lower probability of exceedence (i.e. a higher return period). Note that for most of the strongly impacted rivers, the threshold magnitudes in cumecs (m³ s^-1) are small.
• Mainly in Russia and Canada, Q5v2.1 is lower than Q5v2.0, simply because of the removal of some drier years for Q5v2.0 calculation. In some cases the removed years are a consequence of non-stationary (a clear trend in the reanalysis time series), in others it is only due to some variability in the period with some years of naturally lower flow.
  

A note about trends: Some parts of the world show trends in the GloFAS river discharge reanalysis driven by ERA5, in both GloFAS v2.0 and  GloFAS v2.1. These trends can either be real changes in the climate over the long period (areas becoming drier or wetter over the years), or be related to artificial signals, e.g. related to the inhomogeneities in the observations that were used in the ERA5 reanalysis in producing runoff (especially the proliferation of satellite observations in the period after 1990, etc.). Such trends need to be better identified and understood so that they don't impact the extreme value threshold estimation and their use in defining flood warnings. Work is currently ongoing within the GloFAS team to develop ways to better define and use flood thresholds in GloFAS. Further improved methodology and practice is expected to be implemented in the GloFAS system in 2020.

a.) North America	b.) Europe	c.) Asia
d.) South America	e.) Africa	f.) Australia

Figure 1.: Proportional difference between 5-year thresholds in GloFAS v2.1 (Q5v2.1) and in GloFAS2.0 (Q5v2.0) plotted for river pixels above 1000 km2 (values are defined as (Q5v2.1 - Q5v2.0)/ Q5v2.0).