The last decades have been characterized by a drastic global increase in magnitude and frequency of flood events [1
]. West Africa countries have been affected by a large number of extreme events and have suffered serious flood-related damages [3
]. In particular, the Sahelian area has recorded a strong rain inter-annual variability and an increase in magnitude and frequency of extreme rainfall events [4
], coupled with an increase in flood occurrences [6
]. The amount of precipitation is greater than that observed during the severe drought from 1970 to 1990, but lower than that observed from 1950 to 1960 [8
]. The anomaly, called the “Sahel Paradox”, is that the streamflow in Sahelian rivers is higher compared to those from 1950 to 1960 [9
]. The reason for the higher streamflow is heavily debated in the research community and cannot be entirely explained by a single factor [11
]. The hydrological changes may be caused by the increment of the extreme rainfall events [13
], the changes in land use and land cover [14
], and the rupture of the endorheic basins [17
Structural measures have been solicited by floods with magnitudes greater than every event ever observed in Sahelian rivers, but they have been proven to be insufficient [6
]. Therefore, non-structural measures and a change of perspective in the hydrological analysis are needed to deal with these events [21
]. In recent years, flood early warning systems (EWS) have become a fundamental tool in addressing these needs in order to alert the exposed communities and improve safety [24
]. The earliest information on the forthcoming flood can be generated both from upstream field observations and hydrological models [27
]. The field observations are reliable on the actual river flow but they may generate alerts with an insufficient lead time to prepare adaptive strategies [28
]. For this reason, the employment of outputs from hydrological models is very important in medium and small river basins, where the warning time is not sufficient if based on observed flow [29
]. Sirba, a medium Sahelian river basin shared between Niger and Burkina Faso, perfectly meets these conditions. In the last years, the riverine communities have been affected by frequent and intense flood events [27
]. This has caused enormous damages to the population, whose livelihood is mainly related to family subsistence agriculture.
In such context, this study aims to evaluate and improve the performances of the GloFAS (Global Flood Awareness System) hydrological application of the Copernicus Emergency Management Service in the Sirba River basin. GloFAS is a global probabilistic system providing discharge forecasts all over the globe, with flood information published every day on more than 2000 reporting points [31
]. The model, initially un-calibrated in the 1.0 version, has been recently updated to a 2.0 version that is calibrated on over 1000 flow series [32
]. The evaluation of the model reliability was conducted with continuous, categorical, and skill indices on both GloFAS versions 1.0 and 2.0 [33
]. Unfortunately, due to unreliable meteorological input in the Sahelian area and the hydrological limitations of GloFAS in the study area, the performances have been proven to not to be completely satisfactory [30
]. The weak performances implied that the quality of the forecasts could be improved through an optimization process. In this case, the optimization was conducted from a user point of view rather than directly considering the model parameters. Therefore, the optimization consists in the application of corrective factors to the model outputs [36
]. These correction factors were computed by linear regression models based on the homogeneous periods of the river hydrology [37
Hence, the aim of this research was the quality verification and the optimization of GloFAS results with the purpose of making them available for the Sirba EWS. The work is structured as follows: Section 2
focuses on the study area, hydrological model, and materials and methods adopted. Section 3
describes the results and discusses the significance of the research. Section 4
contains the conclusions and the future perspectives.
3. Results and Discussion
The results jointly considered the four forecasted series (5 days GloFAS 1.0 and 2.0 both in original and optimized format) and compared them with the observed flow series. The analysis timeframe covered the five years (2013–2017) of verification.
The preliminary analysis was conducted on the basic statistical parameters of maximum, minimum, and mean reached during the verification period (Table 6
). The minima displayed that zero wa correctly forecasted by all models. Mean and maxima showed that the raw forecasts heavily underestimated the river discharge—the mean value was less than 1/10, whereas the maximum was less than 1/5 of the observed flow, in both versions of GloFAS. Moreover, G2 values were significantly lower than G1. After the optimization, mean values showed an over-forecasting in OG1 and an under-forecasting in OG2, although the maxima were both over-estimated (Table 6
The flow duration curves confirmed the previous results, leading to some interesting observations: (1) G1 had serious problems in producing zero flow values (even for Q355
); (2) the central part (Q60
) was very similar between G1 and G2; (3) the highest values (Q1
) were quite close in OG1, whereas in OG2 the Q1
is more than double the Q5
; (4) in the interval Q5
, the forecasts were over-forecasted in OG1 and under-forecasted in OG2; and (5) the behavior of OG1 and OG2 was the same from Q80
onwards, even if OG2 was higher than OG1 until Q200
and lower in the final part of the curve (Figure 4
The hydrograph of observed and forecasted flows highlights the ability of all forecasts to capture the annual flow cycle, especially during the dry season (Figure 5
). However, the wet season peaks were widely underestimated in the original versions, whereas the optimized forecasts tended to produce some outlier peaks. The outliers occurred every year, except for 2016 (OG2 in 2013 and 2017, and OG1 in 2014, 2015, and 2017).
A single wet season is shown in Figure 6
in order to better describe the model behavior. The original models predicted values near zero for most of the time period. It is noticeable that in G2 the flow started at the end of August, whereas in G1 it began 15 days before. The optimized versions clearly showed an over-forecasting in OG1 and an under-forecasting in OG2. With regards to the peak values, observed flow and forecasts were quite different: (1) OG2 captured the start and the end of the highest flows (early August to mid-September), whereas OG1 identified the initial peak but overestimated the flow at the end of September; (2) the peaks were quite accurate in OG2, whereas OG1 over-forecasted them; (3) both optimized models properly identified the major peaks (early August and early September), but OG1 over-estimated the peak duration and OG2 under-estimated it.
The 2.0 version of GloFAS did not produce substantial improvements to the forecasts, according to RSR and NSE, and the overall quality even slightly deteriorated. However, some noteworthy enhancements could be reached with the optimization procedure (Table 7
On the basis of performance ratings in Table 8
, G1, G2, and OG1 showed unsatisfactory values, whereas the results obtained by OG2 were sufficiently satisfactory (RSR = 0.72 and NSE = 0.48) [34
]. The optimization improved the performance of RSR (OG1 < G1 and OG2 < G2) and NSE (OG1 > G1 and OG2 > G2) for both model versions. However, it is interesting to note that OG2 performed better than OG1, although G2 achieved worse results than G1 for both RSR and NSE. Therefore, the improvement was significant only for OG2.
The categorical indices displayed different results (Table 9
). BIAS values underlined the under-forecasting or the over-forecasting in terms of threshold exceeding. Therefore, G1, G2, and OG2 under-forecast the yes events, whereas OG1 showed an over-forecasting. The BIAS values exhibited a redundancy (OG1) and a deficiency (G1, G2, and OG2), which suggest that the FAR for OG1, and the POD for G1, G2, and OG2 would not be particularly satisfying. POD and FAR are the fundamental parameters that are analyzed for an EWS application. The best POD performance belonged to OG1, whereas the best FAR performance belonged to G2. However, these values were quite unsatisfactory for all the models because FAR ≥ 60% and POD ≤ 33% were slightly insufficient in activating an EWS mechanism. PC was influenced by the “correct negatives” rather than by the “hits” due to the high amount of values under the yellow threshold. The percentage of “correct negatives” decreased in the optimized models. Therefore, the under-forecasted OG2 appeared to be more accurate than OG1. The quality of forecast in predicting threshold exceeding was assessed by TS. This index considered both “hits” and “false alarms”, thus incorporating the POD and FAR features. It demonstrated that, giving the same weight to POD and FAR, G2 and OG1 were slightly better than G1 and OG2 for an EWS application. TS confirmed the non-optimal behavior observed in POD and FAR and quantified that only 13% (G2 and OG1) and 5–8% (G1 and OG2) of threshold exceeding were correctly identified. HSS considers both the success ratio (1-FAR) and the number of correct random chance forecasts. This skill index is generally used to evaluate rare events [63
]. The best results, such as TS, were reached by G2 (0.21) and OG1 (0.18).
The analyses demonstrated that the GloFAS 1.0 system badly forecasted the flow in the Sahelian area as declared by Alfieri et al. in 2013 [31
]. The Sahelian discharge forecasting complexity is related both to a non-homogeneous watershed response and to several difficulties in correctly predicting the meteorological forcing [35
]. The calibration conducted by Hirpa et al. in 2018 [32
] for GloFAS 2.0 contributed to the improvement of the forecast behavior in terms of categorical indices. On the contrary, the continuous indices were not improved because the calibration was realized before the revision of the Garbey Kourou historical discharge series by using an outdated and unreliable series [30
]. The optimization allowed for the alignment of the forecasted hydrograph shape to the observed flow, even though the flow magnitude was still underestimated.
The continuous indices showed that (1) G1 and G2 are flow predictors less precise than the mean flow, (2) OG1 insufficiently improved the quality of the model, and (3) OG2 obtained results fairly satisfactory according to the classification of Moriasi et al. in 2007 [34
]. Although the performances of forecasts were lower than the literature values for the continuous indices, OG2 proved to be the best solution to forecast the river flow [34
]. The categorical indices demonstrated the over-forecasting of OG1 and the under-forecasting of OG2. Thus, both original and optimized models showed a poor reliability in predicting flood events in the Sirba River Basin. The high number of “false alarms” and the low number of “hits” do not allow for people to be alerted. The forecasts correctly identifying the threshold exceeding were indeed about 10% (13% G2 and OG1, and 5–8% G1 and OG2). Therefore, OG2 can be used to predict the hydrological evolution but not to activate the alert mechanisms.
The analyses underlined the fundamental role of the hazard thresholds—the GloFAS thresholds were less than 1/3 (G1) and 1/10 (G2) of the observed and field-calibrated thresholds [27
]. The categorical indices showed that coincidental events were possible and quite common, although the flow thresholds were completely different. However, it is very important to consider the GloFAS warnings instead of the actual forecasted flow. The Sirba case of study demonstrated that even the maximum forecasted flow did not cause any damage to the riverine communities because it was an ordinary discharge that is overpassed 90 days every year.
The Sahelian hydrological behavior has totally changed in the last decades, generating a high number of floods characterized by unprecedented magnitudes. The change has been caused by the increase of extreme rainfall events coupled with land use changes that generated the rupture of the endoreic basin and the growth of the secondary river network. The Sirba River, one of the major tributaries of Niger River, has been particularly touched by this phenomenon. Flood-related losses have pushed the national departments and the scientific community to develop an early warning system.
Previous studies have already analyzed the hydrology. The joint use of the hydraulic model with the hydrometric observations allows for exposed villages to be alerted a day earlier thanks to hydrometric observations. Therefore, the aim of this research was to evaluate the application of GloFAS (Global Flood Awareness System) in the Sirba River in order to predict the arrival of high flows a few days in advance. The study was conducted on two GloFAS systems: the original GloFAS 1.0 and the newest 2.0 version. The analysis evaluated and optimized the re-forecasts of a 10 year period (2008–2017). The 5 days forecasts were chosen in order to have adequate time to activate the field adaptive measures. The optimization was conducted from the user point of view by correcting the model outputs instead of the model parameters. The datasets were split into 5 years for training and 5 years for validation. The computation of the correction factors was performed through linear regression between observed and forecasted flow. The validation considered continuous and categorical indices.
The results showed the poor reliability of both version 1.0 and 2.0 for the original forecasts. These forecasts, as flow predictors, were less accurate than the mean flow. Although the calibration for version 2.0 conducted by GloFAS developers improved the EWS skills, the flow deficit increased because the calibration was based on an outdated and unreliable flow series of the Garbey Kourou hydrometer. The optimization procedures produced a substantial improvement in forecast accuracy. The continuous indices of optimized GloFAS 2.0 were quite satisfactory with regard to flow prediction, whereas the improvements were not conspicuous in the 1.0 optimized model. Therefore, the enhancement of GloFAS 2.0 demonstrated the importance of the optimization in regulating the shape of the forecasted hydrograph rather than in adjusting its intensity. The reliability of the flow peak forecasts, measured by the categorical indices, was low for both the original and optimized models, as the percentage of correctly predicted floods was approximately 10%. The results also illustrated the level of hazard threshold on which to develop a hydrological model in order to correctly quantify the river discharges and to not only issue warnings. This outcome could be used for a new version of the GloFAS system.
However, the optimized GloFAS 2.0 will be used in the EWS platform for the Sirba River. This application will be useful in providing reliable information on the evolution of river hydrology, but less appropriate to send alerts to the exposed populations. In order to guarantee the reliability of the EWS platform, the alerts will only be sent using the in situ measurements.
Future work will involve an enhanced collaboration with hydrological model developers and the implementation of the Sirba EWS platform with the new updated versions of GloFAS. Model calibration on the observed flow series or the utilization of a regional hydrological model could improve peak forecasting. These studies will allow the use of hydrological models for the EWS, which is currently conducted with in situ observations only.