398
A. Amani et al.
performance of the model and a more comprehensive analysis is required to understand its reliability. Nonetheless, the evaluation provides an initial estimate of the
potential of the system. The GEFS provides 20 ensemble forecast members which
are aggregated to daily, downscaled and bias-corrected to the LCB domain at 0.05°
resolution. Bias correction is against the historic data in the Lake Chad Monitor for
the overlap period. Forecasts are carried out for lead times from 6 to 0 days.
Figure 5 shows an example of the 6-day lead forecasts of precipitation for 2 May
2018: the forecast was initialized on 26 April 2018. The 20 ensemble members are
shown and compared to the observed data for the forecast date. Note the spread in
the ensembles, which generally encompass the spatial footprint of the observed data,
but tend to underestimate it overall.
Figure 6 shows a summary of the accuracy in terms of the spatial correlation of
the forecasts across the three example dates and for each lead time. Correlation is
calculated across the wetter part of the basin up to latitude 15
o N to avoid biasing the
results to the strong wet-dry meridional gradient of precipitation. The lines represent
the mean of the ensemble and the error bars are ± one standard deviation of the
ensemble. Precipitation and runoff correlations are modest (<0.4), which is expected
given the difficulty in forecasting convective systems. Correlation tends to reduce
with increasing lead time, as is also expected. The soil moisture correlations are much
higher given its persistence, with much lower uncertainty across ensemble members.
Figure 7 shows an example of the basin average forecast for the three variables
with a 6-day lead, highlighting the improved accuracy when averaging over larger
scales, even with this somewhat long lead time. The uncertainty for the precipitation is however quite large, with some ensemble members predicting high and early
precipitation. The translation of this uncertainty into total runoff and soil moisture is
dampened and their relative accuracy is higher, although with a tendency to underestimate the observed data; if this is found to be systematic, it can be bias-corrected
as a post-processing step.
A. Amani et al.
performance of the model and a more comprehensive analysis is required to understand its reliability. Nonetheless, the evaluation provides an initial estimate of the
potential of the system. The GEFS provides 20 ensemble forecast members which
are aggregated to daily, downscaled and bias-corrected to the LCB domain at 0.05°
resolution. Bias correction is against the historic data in the Lake Chad Monitor for
the overlap period. Forecasts are carried out for lead times from 6 to 0 days.
Figure 5 shows an example of the 6-day lead forecasts of precipitation for 2 May
2018: the forecast was initialized on 26 April 2018. The 20 ensemble members are
shown and compared to the observed data for the forecast date. Note the spread in
the ensembles, which generally encompass the spatial footprint of the observed data,
but tend to underestimate it overall.
Figure 6 shows a summary of the accuracy in terms of the spatial correlation of
the forecasts across the three example dates and for each lead time. Correlation is
calculated across the wetter part of the basin up to latitude 15
o N to avoid biasing the
results to the strong wet-dry meridional gradient of precipitation. The lines represent
the mean of the ensemble and the error bars are ± one standard deviation of the
ensemble. Precipitation and runoff correlations are modest (<0.4), which is expected
given the difficulty in forecasting convective systems. Correlation tends to reduce
with increasing lead time, as is also expected. The soil moisture correlations are much
higher given its persistence, with much lower uncertainty across ensemble members.
Figure 7 shows an example of the basin average forecast for the three variables
with a 6-day lead, highlighting the improved accuracy when averaging over larger
scales, even with this somewhat long lead time. The uncertainty for the precipitation is however quite large, with some ensemble members predicting high and early
precipitation. The translation of this uncertainty into total runoff and soil moisture is
dampened and their relative accuracy is higher, although with a tendency to underestimate the observed data; if this is found to be systematic, it can be bias-corrected
as a post-processing step.