GloFAS 3.1 hydrological performance evaluation was conducted over the period for the period 1979-2019, with specific period of the year defined for each catchment based on the maximum observed discharge in the daily climate mean time series, and for GloFAS verifications stations selected from a pool of stations available to the GloFAS team. For the methodology details please see the section GloFAS hydrological performance.
This section documents the GloFAS v3.1 hydrological model performance assessment, based on GloFAS v3.1 historical reanalysis available at https://cds.climate.copernicus.eu/cdsapp#!/dataset/cems-glofas-historical?tab=overview.
Catchments
GloFAS 3.1. hydrological performance evaluation was conducted on 1532 verification stations (Figure 1), including 982 catchments used in GloFAS v3.1 calibration, of which 795 catchments were in both GloFAS v3.1 and v2.1 calibrations. An additional 351 catchments where used in this verification which were in none of the GloFAS calibration exercises (but some of these could be sub-catchments of larger calibration catchments).
Figure 1: GloFAS verification stations (April 2021) and observational record length.
General hydrological performance
Table 1 provides an overview of all GloFAS v3.1 hydrological performance scores obtained for all the 1532 verification stations, each verified for their respective main flood season only. while Figure 2 shows each of the five main scores at the 1532 station locations.
GloFAS v3.1 has a dominantly positive bias (median pbias of 0.25). In the middle and higher latitudes of the Northern Hemisphere, pbias is generally neutral (balanced around 0), but the bias is dominantly positive in tropical areas and in central parts of the USA. In these areas many catchments highlight verification-period-averaged river discharge which is three or more times larger in the simulation than in the observations (dark blue areas with pbias>2; seen in 15% of catchments).
Correlation is high in the majority of the catchments (median pcorr of 0.70), with lowest correlations mainly present in some central parts of North America, some areas in Africa, catchments in India and Australia.
The GloFAS v3.1 simulation has lower variability than the observations (median of -0.12, and the median of the absolute variability errors of 0.18). It can be noted that larger negative variability errors show a similar pattern to where low correlation tend to occur.
The KGE' median value is 0.47, with a clear link between low KGE' and pbias errors, so lower KGE' over areas where pbias tend to show a larger positive error.
The timing error is less than equal to 1 days (-1, 0 or +1 day) in about 50% of the catchments and reaches about 9-10 days (either positive or negative) only in about 10% of the catchments (6% for negative and 4% for positive errors). No timing error is identified in about 20% of the catchments, with errors larger than -+3 days found in about 35% of catchments.
In Table 1, the global mean of the scores are also provided, based on the 1532 analysed catchments. For pbias and KGE, the mean is rather skewed, as the tail values of the very large positive pbias and very low negative KGE' (occurring only at 1 or two catchments) are not capped, and thus the mean is distorted somewhat, being disproportionally too high for pbias and too low for KGE'.
It is also important to point out, that the flood-season-based verification scores are a little bit lower than the scores based on the whole year. The whole year based score medians are (equivalent to the red line in Table 1) as follows: 0.27, 0.29, -0.09, 0.15, 0.80, 0.54, 1, 1.
Table 1. GloFAS v3.1 river discharge reanalysis score distribution and mean based on 1532 global catchments.
Figure 2: Hydrological performance scores for GloFAS v3.1.
Catchments north of 60N latitude have dominantly high KGE', mainly above 0.5, whilst both high and very low KGE' values are found in lower latitudes, especially in the 0-20 degree band in the tropics (Figure 3, left hand side). There is less of a strong influence between KGE' and catchment area, but larger catchments are more likely to have higher KGE' - catchments over 350.000 km2 (over ~70 in Figure 3, right hand side) tend to have KGE' within 0.5 and 1 (Figure 3) with a smaller proportion of exceptions. It is good to note, that there is a clear step change in Figure 3 (right hand side), with lot more catchment from 10.000 km2. This is because in the early years of GloFAS, the GRDC stations were only added to GloFAS from 10.000 km2 catchment size. In the last few years, this minimum catchment size was lowered and now observation data is used for also smaller catchments, usually above 1000 km2.
Figure 3: GloFAS v3.1 KGE' score distribution according to the latitude of the stations (left/top) and according to the upstream area size (right/bottom). The upstream area is provided as the cube root of the area on the x-axis (~22 is the equivalent to 10000 km2 while 100 is for 1.000.000 km2). The colour of the dots represent the latitude (dark blue is equatorial, green is mid-latitudes and yellow is high latitudes), while the dot size the upstream area.
Performance in calibrated vs non calibrated stations
Hydrological performance scores are notably lower for stations not used for the v3.1 model calibration (Tables 2, Cal-No columns), vs stations which were calibrated. This is generally expected, however, there are still some interesting aspects to highlight:
- The correlation decrease is relatively moderate, with median pcorr of 0.59 for stations not used in the calibrations compared with 0.74 when only calibration catchments are analysed
- Very significant worsening of the pbias and KGE' errors (strongly linked to pbias), with median absolute pbias of 1.14 vs 0.13 and KGE' of -0.34 vs 0.60 for non calibration vs calibration catchments, respectively
- Some worsening of timing errors in the non-calibration catchments, with the absolute timing error (abstiming) being 5.19 vs 3.41 days as global average. Also, the +-1 day error proportion increases from 45% to 50% of the catchments, while the +-5 days errors increase from 74% to 84% in non-calibration stations vs calibrated stations.
One hypothesis for the poorer hydrological performance is the use of default parameter maps in non calibrated catchments, which are probably not suited for some areas of the world, such as the tropics (e.g. Africa and India), resulting in large positive pbias and hence low KGE' (Figure 4).
Table 2. GloFAS v3.1 scores for catchments included and not included in the calibration.
Figure 4: GloFAS v3.1 hydrological performance scores for calibration catchments (left column or 1st maps) and catchments not included in the calibration (right column or 2nd maps).