Summary
GloFAS 3.1 hydrological modelling forwent a major change in GloFAS v3.1, with a completly different hydrological model configuration, moving away from a NWP-hydrological routing (with a calibration of the routing component for v2.0) to a global hydrological model set-up fully calibrated for v3.1. This means, there are possible differences in the way river discharge is simulated, which are investigated in this section. The methods and scored used are described in GloFAS hydrological performance.
For a comparison as fair as possible, the analysis was undertaken on two groups of stations:
- Stations were used as calibration points for GloFAS v2 and GloFAS v3
- Stations not included as calibration points in neither GloFAS v2 nor GloFAS v3
In the following, we summarise the main results score by score, with details presented in Table 1 and 2 with the score distribution in the calibration and the non-calibration station groups. They are complemented with graphs and maps for both groups and for each hydrological performance score.
Bias (pbias)
GloFAS v3.1 generally shows a higher river discharge and thus pbias than GloFAS v2.1 (Table 1-2 and Figure 1) for both groups of stations.
For calibration stations, river discharge is generally 43% higher (as a global average of all analysed catchments) in v3.1 than in v2.1 (i.e. pbias difference has a median of +0.43). Overall, pbias is improved for GloFAS v3.1 for 60% of the stations (based on the abspbias score). The increase of the amount of river discharge helps decreasing the absolute pbias error for the majority of the catchments in the Northern Hemisphere in the calibrated station group (Figure 2). Exception is the central parts of North America, and areas in the tropics, e.g. some areas in South America, large parts of Africa and India, where pbias is worse in v3.1 than in v2.1; in those regions the generally higher water amount is detrimental to the simulation quality (Figure 2). GloFAS v2.1 was generally lacking water in the rivers, appearing to show negative pbias in about 76% of the catchments, while in v3.1 it is the opposite with 71% of them (pbias being higher than 0) having too much water in the simulation compared with the river discharge observations.
For non-calibration stations, pbias is improved in GloFAS v3.1 compared with v2.1 in 34% of the catchments (the absolute pbias error decreases). Globally, GloFAS v2.1 had a median pbias close to 0 (about 50% of the stations with positive/ negative bias). However, pbias in GloFAS v3.1 is generally very large, with 50% of the stations showing at least twice as high average discharge than the average observation (Table 2).
Variability (var)
GloFAS v3.1 generally shows a decrease in the variability error compared with GloFAS v2.1 (Table 1-2 and Figure 3-4). The improvement is more widespread in the Northern Hemisphere, with exception of the central parts of North America and central Amazonia.
The absolute variability error (absvar) decreases in 61% of the catchments in the calibration station group and 39% of the non-calibration group.
Correlation (pcorr)
GloFAS v3.1 sees a largely improved correlation compared with GloFAS v2.1, with closer match between simulated and observed time series on average (Table 1-2 and Figure 5). The improvements are more present in much of the Northern Hemisphere, with quite few of the stations depicting an increase of at least 0.4 correlation (dark blue catchments). The tropics, on the other hand, shows more mixed behaviour with clearly less advantage of v3.1 over v2.1.
In total, 84% of the catchments show improved correlation compared with v2.1 amongst the calibration stations and 71% for non-calibration stations.
Modified Kling Gupta Efficiency (KGE')
GloFAS v3.1 displays a better KGE' compared with GloFAS v2.1 in much of the Northern Hemisphere and central area in Amazonia (Table 1-2 and Figure 6). However, the central parts in North America, most of the catchments in Africa and quite a few of them in South America have lower KGE' in v3.1 than in v2.1. The geographical distribution of the KGE' errors closely resembles the absolute pbias error difference patterns, confirming that the dominant issue in the simulations limiting hydrological modelling performance, is the underlying biases.
In total, 71% of the catchments in the calibration group shows higher KGE' scores in v3.1 than in v2.1. The KGE' in the non-calibration stations shows noticeably worse performance, similarly to the case with the other metrics above, with the majority of the catchments in the tropics having lower KGE' in much of Africa and also in India in v3.1. In total, only 38% of the catchments show improvement in v3.1 amongst the non-calibrated catchments.
Timing error (timing)
Performance in GloFAS v3.1 timing compared with v2.1 is associated with a rather mixed picture geographically (Figure 7-8). Generally, the absolute timing error (abstiming) decreases (improved simulation) in the majority of the catchments in the middle and higher latitudes in North America, Europe and Asia. However, larger timing errors (increase in abstiming) are found for v3.1 in the central area of the USA, and quite few catchments in the tropics.
In total, GloFAS v3.1 timing is improved for about 50% of the calibration stations and is deteriorated for 32% of them (in the remaining part the timing error does not change). In the non-calibration group, timing is improved in about 45% of the stations, and is deteriorated in 34% of the group.
Score tables
Table 1. Distribution of hydrological performance scores for GloFAS v3.1 and v2.1 and in their score difference (sv3.1-sv2.1) computed with stations that were included in the model calibrations
Table 2. Distribution of hydrological performance scores for GloFAS v3.1 and v2.1 and in their score difference (sv3.1-sv2.1) computed with stations that were NOT included in the model calibrations.
Score graphics
Pbias scores
Figure 1. Pbias score maps for the calibration and non-calibration station (top row) and the corresponding distributions in v3.1 and v2.1, and the distribution of the difference pbias values (left: calibration stations, right: non-calibration stations).
Abspbias scores
Figure 2. Abspbias score maps for the calibration and non-calibration station (top row) and the corresponding distributions in v3.1 and v2.1, and the distribution of the difference abspbias values (1st two plots for calibration, 2nd two plots for non-calibration stations).
Var scores
Figure 3. Var score maps for the calibration and non-calibration station (top row) and the corresponding distributions in v3.1 and v2.1, and the distribution of the difference var values (1st two plots for calibration, 2nd two plots for non-calibration stations).
Absvar scores
Figure 4. Absvar score maps for the calibration and non-calibration station (top row) and the corresponding distributions in v3.1 and v2.1, and the distribution of the difference absvar values (1st two plots for calibration, 2nd two plots for non-calibration stations).
Pcorr scores
Figure 5. Pcorr score maps for the calibration and non-calibration station (top row) and the corresponding distributions in v3.1 and v2.1, and the distribution of the difference pcorr values (1st two plots for calibration, 2nd two plots for non-calibration stations).
KGE' scores
Figure 6. KGE' score maps for the calibration and non-calibration station (top row) and the corresponding distributions in v3.1 and v2.1, and the distribution of the difference KGE' values (1st two plots for calibration, 2nd two plots for non-calibration stations).
Timing scores
Figure 7. Timing score maps for the calibration and non-calibration station (top row) and the corresponding distributions in v3.1 and v2.1, and the distribution of the difference timing values (1st two plots for calibration, 2nd two plots for non-calibration stations).
Abstiming scores
Figure 8. Abstiming score maps for the calibration and non-calibration station (top row) and the corresponding distributions in v3.1 and v2.1, and the distribution of the difference abstimings values (1st two plots for calibration, 2nd two plots for non-calibration stations).