A Flood Forecasting Method in the Francolí River Basin (Spain) Using a Distributed Hydrological Model and an Analog-Based Precipitation Forecast

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript proposes a probabilistic flood forecasting analysis for the Francolí River Basin in Catalonia, Spain, in response to recent flood events in the region. The methodology involves calibrating the distributed hydrological model Real-time Interactive Basin Simulator (RIBS) for a series of flood events.

I have the following major observations:

In such flood risk analyses the most important aspect is the future floods related to the annual exceedance probabilities of interest. The analysis is done exclusively by modeling recorded events but does not make a probabilistic extrapolation of them to estimate the magnitude of events with long mean return periods (flood protection works are designed for such probabilistic values). For small basins it is better to do a flood frequency analysis ( https://doi.org/10.3390/w17121799 , https://doi.org/10.3390/w17101539 ) and make decisions to protect the objectives based on these forecast values because they are more conservative and characterized by fewer uncertainties. The authors recognize the uncertainties inherent in hydrological models and parameter estimates, but do not make a rigorous quantification of them.

The calibration of 5 events is too superficial. The number of events is too small. Although five flood events are used for calibration, the distribution of these events (in terms of category - ordinary, extraordinary, catastrophic) and their frequency could be discussed in more detail to justify why they are sufficient for a robust calibration.

The results show that the rainfall forecasts based on analogues and extreme analogues show lower rainfall than the actual ones in the four events analyzed, although with variability. This systematic underestimation should be analyzed in more depth, including its implications for flow forecasts and potential correction methods.

In the results section, it can be observed that some simulations (2010 and 2015 events) have negative NSE values, indicating that the model is a poor predictor, even worse than the average of the observations. Although the authors note that "visually, there are examples where better calibration results are obtained, as in the case of the flood events of 2013, 2014 and 2018", and select Calibration 2 as the best, a more thorough discussion of the poor performance for certain events and its implications for the overall reliability of the model would be valuable.

Author Response

We thank the anonymous reviewer for their feedback, questions and comments that have helped improve the quality of the manuscript. Below these lines is a point-by-point answer to all the reviewer's questions and comments. All changes made in the original manuscript are highlighted in red.

The manuscript proposes a probabilistic flood forecasting analysis for the Francolí River Basin in Catalonia, Spain, in response to recent flood events in the region. The methodology involves calibrating the distributed hydrological model Real-time Interactive Basin Simulator (RIBS) for a series of flood events.

I have the following major observations:

1. In such flood risk analyses the most important aspect is the future floods related to the annual exceedance probabilities of interest. The analysis is done exclusively by modeling recorded events but does not make a probabilistic extrapolation of them to estimate the magnitude of events with long mean return periods (flood protection works are designed for such probabilistic values). For small basins it is better to do a flood frequency analysis ( https://doi.org/10.3390/w17121799 , https://doi.org/10.3390/w17101539 ) and make decisions to protect the objectives based on these forecast values because they are more conservative and characterized by fewer uncertainties. The authors recognize the uncertainties inherent in hydrological models and parameter estimates, but do not make a rigorous quantification of them.

We thank the reviewer for this important observation. Indeed, for flood risk analysis and estimation of future flood magnitudes for large return periods, it is important to calculate annual exceedance probabilities and extrapolate past events into future scenarios for designing flood protection measures. However, this paper focuses on real-time forecasting in a medium size mediterranean basin, using analog based precipitations predictions and a distributed hydrological model to support decision making process (earlier floods alerts) in short-term events, rather than estimating high return period floods events that are useful in the design of flood protection measures. Therefore, the statistical analysis falls outside the scope of this paper. We have added a sentence clarifying this aspect in the introductory section of the revised version of the manuscript. [lines 190-194]

Additionally, regarding the uncertainties inherent in hydrological models and parameter estimates, we acknowledge that the reviewers' proposal is both interesting and relevant. However, this aspect falls beyond the scope of the current paper. Therefore, we recognize this limitation, and we have added a statement in the conclusion section emphasizing its significance and proposing to study it in future research.

An extensive sensitivity analysis of the RIBS hydrological model parameters can be found in https://doi.org/10.1080/02626667.2011.610322 so we have referenced this work to provide further details in that sense. Nevertheless, to illustrate how sensitive the flood forecast is to changes in parameters in the Francoli river catchment, a study has been carried out in the event of 14 November 2005 (for the case with 36 h of real warming). Therefore, Section 4.1.2 has been added.

2. The calibration of 5 events is too superficial. The number of events is too small. Although five flood events are used for calibration, the distribution of these events (in terms of category - ordinary, extraordinary, catastrophic) and their frequency could be discussed in more detail to justify why they are sufficient for a robust calibration.

We acknowledge the reviewer’s concern regarding the limited number of calibration events. The study basin is located in a Mediterranean climate zone, characterized by highly variable precipitation patterns, where flash floods are predominantly driven by short-duration, high-intensity convective storms. These events are infrequent, highly localized, and often underrepresented in long-term records. Therefore, a usual problem in the Mediterranean area is the lack of information about the most catastrophic and infrequent flood events in the systematic flood series recorded at gauging stations.

Due to the short length and fragmentation of historical streamflow records at the available gauging stations, the five flood events used in this study constitute the most complete set of documented events with sufficient hydrometeorological and discharge data needed for calibration. Each selected event was thoroughly analyzed and spans a representative range of observed flood magnitudes, including ordinary and extraordinary events. Unfortunately, no catastrophic events with reliable discharge data were recorded during the gauging period.

While the number of events is limited, this is a structural limitation of Mediterranean basins with ephemeral or intermittent flow regimes. The five events were selected to maximize variability in rainfall intensity, antecedent conditions, and spatial rainfall distribution, thus enabling a calibration that reflects the basin’s dominant hydrological response mechanisms.

We have included a paragraph in Section 2 [lines 213-227] and in Section 3.2 [lines 313-318] with a more in-depth explanation on this topic, backed up by some references in the literature in the revised version of the manuscript.

3. The results show that the rainfall forecasts based on analogues and extreme analogues show lower rainfall than the actual ones in the four events analyzed, although with variability. This systematic underestimation should be analyzed in more depth, including its implications for flow forecasts and potential correction methods.

We thank the reviewer for this remark. Indeed, the behavior the reviewer is describing is a well-known drawback of this technique. The catchment used as a case study is located in the Mediterranean region, where floods are usually driven by high localised storms. However, the rainfall maps of previous storms used in the analog-based methodology have a spatial resolution that cannot be enough to capture the spatial variability of such storms. This limitation of the analog-based method is stated in the paper in Section 3.3 [lines 388-391].

 

4. In the results section, it can be observed that some simulations (2010 and 2015 events) have negative NSE values, indicating that the model is a poor predictor, even worse than the average of the observations. Although the authors note that "visually, there are examples where better calibration results are obtained, as in the case of the flood events of 2013, 2014 and 2018", and select Calibration 2 as the best, a more thorough discussion of the poor performance for certain events and its implications for the overall reliability of the model would be valuable.

We thank the reviewer for the suggestion. As explained in previous observation, the special conditions of the catchment, characterized by highly variable rainfall patterns with flash floods with short duration of the storm and convective precipitation, make it difficult to accurately reproduce certain event such as those of the years 2010 and 2015.

In this context, the presence of negative NSE values reflects cases where the simulated hydrographs fail to replicate both the timing and magnitude of peak flows. This behavior is characteristic of NSE's sensitivity to peak errors, especially when the observed hydrograph is highly skewed or dominated by a sharp, narrow peak.

However, these negative values should not be interpreted as a general failure of the modeling framework. Instead, they highlight the localized limitations in capturing rapid-response dynamics, often driven by uncertainties in rainfall intensity, spatial variability, or antecedent soil moisture, factors that can disproportionately affect peak representation in event-based simulations.

It is important to note that across the broader dataset, most simulations yield positive and acceptable NSE values, with many exceeding 0.75 and RMSE values falling within a reasonable range. This demonstrates that the model is consistently robust for the majority of events and parameter combinations. Furthermore, simulations with high NSE (e.g., 15 October 2018) show that the model is fully capable of achieving both temporal and volumetric accuracy when boundary conditions are well-represented.

In summary, the few negative NSE values are diagnostic outliers that reveal peak misalignment issues rather than overall model inadequacy. These cases underscore the need for enhanced input data resolution or event-specific calibration, without undermining the broader validity and predictive skill of the model across diverse flood events.

Therefore, the NSE negative values in those events reflect the difficulty for the model to reproduce the magnitude, shape or time of some of these rapid and intense flood events. We have rewritten the corresponding section (4.1 and 4.1.1) of the results to address the reviewer’s concerns.

Reviewer 2 Report

Comments and Suggestions for Authors

This paper shows how to predict floods in the Francolí River basin in Spain by using a distributed hydrological model (RIBS) and an analog-based precipitation forecasting method. The goal of the project is to create a viable technology for early warning of floods in tiny Mediterranean watersheds, which can respond quickly to changes in water levels. The best things about this study are that it combines atmospheric and hydrological modelling, uses a genuine flood event (2019) as an example, and shows how well the models work by running simulations of past events.

Major comments 1. Model validation only works for one flood occurrence. The 2019 flood is important, but it's not clear how far the results may be applied. Using additional occurrences or made-up circumstances to validate the statements would make them stronger.
2. There is no analysis of uncertainty. There isn't enough discussion or quantification of the uncertainties in both the analogue precipitation forecast and the hydrological model output (for example, because of parameter calibration or the quality of the input data).
3. There has to be greater explanation for the choice of analogue dates and selection criteria. Are the predictors and criteria used to choose analogues based on previous research or real-world testing? This has to be made clearer.
4. There is no talk about how sensitive the parameters of the hydrological model are. It would be helpful to know how sensitive the flood forecast is to changes in parameters, because the RIBS model was set up on prior experiences.
5. The operational practicality (such how long it takes to compute, if real-time data is available, and whether automatic integration is possible) is not discussed. These things should be briefly described for a technique that is meant to be used in the field.
6. It would be better if the references included more recent research on using analogue approaches to predict floods and ensemble rainfall predictions for small basins.

Minor comments 1. Lines 70 to 90 The explanation of how the analogue forecasting method works is too short. Please explain how you choose the analogues, such as the weather variables you utilised, the similarity metrics, and the time periods.
2. Figure 4 shows the group of analogue precipitation projections, but it doesn't make it evident how the group is spread apart. Include a discussion of ensemble variability and what that means for the uncertainty of flood predictions.
3. Lines 100–120 The description of the RIBS model doesn't go into enough depth about how to calibrate and validate it. Please tell me more about the events that were used, the performance criteria, and whether any cross-validation was done.
4. Line 135 The model is validated using a single event. This is known, but the restriction should be made more apparent. You might want to think about using simulated rainfall scenarios or longer-term hindcasting in future work.
5. Line 185 There is no indication of lead time or how long before the prognosis was made. This is important for figuring out how useful the procedure is.
6. In Table 1, please include a column that shows the sources of the data or the resolution for each input dataset (e.g., land use, precipitation, topography).
7. Lines 210–220 It could be a good idea to talk about how this method could be used in other Mediterranean catchments or in other areas.  

Author Response

We thank the anonymous reviewer for their feedback, questions and comments that have helped improve the quality of the manuscript. Below these lines is a point-by-point answer to all the reviewer's questions and comments. All changes made in the original manuscript are highlighted in red.

 

This paper shows how to predict floods in the Francolí River basin in Spain by using a distributed hydrological model (RIBS) and an analog-based precipitation forecasting method. The goal of the project is to create a viable technology for early warning of floods in tiny Mediterranean watersheds, which can respond quickly to changes in water levels. The best things about this study are that it combines atmospheric and hydrological modelling, uses a genuine flood event (2019) as an example, and shows how well the models work by running simulations of past events.

Major comments

1. Model validation only works for one flood occurrence. The 2019 flood is important, but it's not clear how far the results may be applied. Using additional occurrences or made-up circumstances to validate the statements would make them stronger.

We thank the reviewer for their observations and comments, but there seems to be some kind of error since there is no 2019 event included in this work. This is because for the EA028 gauging station there was no recorded data for that event.

The calibration and validation were made in five floods events (section 3.2 and 4.1), using the objective functions to obtain the optimal RIBS model parameters for the Francoli river catchment. We acknowledge the limitations of using only five events, but the limited data available and the special conditions in the catchment made it difficult to perform a better analysis. This is due to the location of the basin, the Mediterranean area is characterized by highly variable precipitation patterns, where the occurrence of flash floods (very intense convective rainfall over a short period of time) makes it difficult to accurately record these flooding events. This is because events are rare and frequently underestimated in long-term historic records. Therefore, this study selected the events where sufficient rainfall and streamflow data was available. The episodes were chosen to maximize variability in precipitation intensity, antecedent conditions, and spatial distribution of precipitation. The aim was to calibrate the model in a way that most accurately represents the dominant hydrological response mechanisms of the basin. We have included a paragraph in the revised manuscript in Section 2 [lines 213-227] and in Section 3.2 [lines 313-318] explaining this issue in detail.

2. There is no analysis of uncertainty. There isn't enough discussion or quantification of the uncertainties in both the analogue precipitation forecast and the hydrological model output (for example, because of parameter calibration or the quality of the input data).

We acknowledge the reviewer’s proposal. In the paper, we have considered the uncertainty in probabilistic flood forecasts by using a set of analog and extreme-analog rainfall forecasts. Therefore, the forecast is not deterministic, but probabilistic supplying a confidence interval for the discharge in each time step.  The uncertainty of the hydrological model in the simulations has not been included in the results.  However, an extensive sensitivity analysis of the RIBS hydrological model parameters can be found in https://doi.org/10.1080/02626667.2011.610322, so we have referenced this work to provide further details in that sense.

In addition, we consider that a more in-depth analysis of uncertainty is beyond the scope of this paper and will be carried out in future research. We have added a paragraph to the conclusion section explaining this issue.

3. There has to be greater explanation for the choice of analogue dates and selection criteria. Are the predictors and criteria used to choose analogues based on previous research or real-world testing? This has to be made clearer.

Following the reviewer’s request, in section 3.3, we have given a more detailed explanation on the choice of analogs and selection criteria, referencing previous works where the methodology is fully explained.

4. There is no talk about how sensitive the parameters of the hydrological model are. It would be helpful to know how sensitive the flood forecast is to changes in parameters, because the RIBS model was set up on prior experiences.

We thank the reviewer for highlighting this important point. We acknowledge that we have not included an explicit sensitive analysis for the parameters of the RIBS hydrological model in the manuscript. Instead, in the calibration section we tested different parameters sets in five different events, where for each parameter combination, a different hydrograph response was obtained in the same studied episode. This brief analysis of sensitivity reveals how the flood response changes for different parameters sets. It can be consulted in the results of the calibration section 4.1, for both the calibration of each episode (in Table 4 and Figures 5, S1, S2, S3, and S4) and for the six different combinations of parameters tested in the five episodes considered (Table 6 and Figures 6, S5, S6, S7 and S8). Additionally, to illustrate the scale of parameters modifications, Table 3 has been added in the methodology section 3.2. In this table, the ranges of each RIBS parameter are displayed.

An extensive sensitivity analysis of the RIBS hydrological model parameters can be found in https://doi.org/10.1080/02626667.2011.610322 so we have referenced this work to provide further details in that sense.

Nevertheless, to illustrate how sensitive the flood forecast is to changes in parameters in the Francoli river catchment, a study has been carried out in the event of 14 November 2005 (for the case with 36 h of real warming). Therefore, Section 4.1.2 has been added.

5. The operational practicality (such how long it takes to compute, if real-time data is available, and whether automatic integration is possible) is not discussed. These things should be briefly described for a technique that is meant to be used in the field.

We thank the reviewer for the observation. In practice, the method would work as follows in a real flood event. Some hours or days previous to the flood event, we would have available forecasts about the meteorological conditions expected in the next days or hours. More specifically, information about 500 hPa and 1 000 hPa geopotential height fields supplied by GFS will be considered. With such information, the analog-based method to obtain the possible rainfall fields in the future can be applied in few minutes. Finally, the RIBS model will be used to simulate the catchment response with such rainfall fields as input data in few minutes. Therefore, the probabilistic forecast result can be delivered in around 15 minutes.  We have included a new subsection 5.3 where we added new information in this regard.

6. It would be better if the references included more recent research on using analogue approaches to predict floods and ensemble rainfall predictions for small basins.

Following the reviewer’s suggestion we have included some recent references on the topic in the introduction section.

 

Minor comments

Concerning the minor comments from the reviewer, there seems to be some kind of misunderstanding regarding line numbers because the line numbers referred to by the reviewer do not correspond to the alluded points. We have tried our best to understand the remarks, but some may not be as accurate as expected.

1. Lines 70 to 90 The explanation of how the analogue forecasting method works is too short. Please explain how you choose the analogues, such as the weather variables you utilised, the similarity metrics, and the time periods.

Regarding the reviewer’s concerns about how the analogs method works, an extended explanation has been added in section 3.3 of the reviewed manuscript.

2. Figure 4 shows the group of analogue precipitation projections, but it doesn't make it evident how the group is spread apart. Include a discussion of ensemble variability and what that means for the uncertainty of flood predictions.

We thank the reviewer for the observation. Figure 4 (Following the revisions to the manuscript, it is now Figure 5) refers to calibration of 2018 event, where the four hydrographs that match better the observed one are represented. For this 2018 event, visually (substantiated by the RMSE and NSE similar values), there is little spread, as the four calibration hydrographs are very similar. We have included an explanatory text below the figure. In addition, for greater clarity regarding the six combinations of parameters evaluated and used to select the model parameters, Table 5 has been added in the revised version of the manuscript.

3. Lines 100–120 The description of the RIBS model doesn't go into enough depth about how to calibrate and validate it. Please tell me more about the events that were used, the performance criteria, and whether any cross-validation was done.

We thank the reviewer for the observation. In the paper, it is explained the calibration and validation process made for the Francolí river catchment. Some additional references were provided that address the operation, calibration and validation process of the RIBS hydrological model (https://doi.org/10.1080/02626667.2011.610322)

In the calibration process, five events with major flood events in the past were identified and used as calibration events (Table 1). For each event, optimal model parameters were selected to achieve a hydrograph that closely matched the observed one (Table 4). After obtaining the best parameters for each of the five events (with a range between 1 to 4 simulations per event), six different parameter combinations were evaluated, listed in table 5. Then, was evaluated how each of these combinations performed across the five events considered in the calibration (Table 6). Finally, the parameter combination that achieved in global the better NSE and RMSE values was selected for the simulations with the analogs in other five different events (Table 7). No cross-validation was carried out; we have used all the events as part of the training process.

4. Line 135 The model is validated using a single event. This is known, but the restriction should be made more apparent. You might want to think about using simulated rainfall scenarios or longer-term hindcasting in future work.

We appreciate the reviewer's observations and acknowledge the limitations of the calibration and validation process. As explained above, the calibration and validation process were made for the five events of sections 3.2 and 4.1. Then, other episodes had been used to simulate analog events. In addition, it has not been possible to use more episodes for the calibration due to the limited number of properly recorded historical flood events that possess sufficient hydrometeorological and discharge data for calibration. Nevertheless, in the future we intend to improve the calibration and the analog methodology to achieve better results. A paragraph has been added to clarify the limitation of calibration episodes in Section 3.2.

5. Line 185 There is no indication of lead time or how long before the prognosis was made. This is important for figuring out how useful the procedure is.

The catchment used as a case study is subject to flash floods with short response times of few hours. Therefore, flood forecasts are needed some hours before the beginning of a given flood event. In this catchment, flood forecasts generated with observed precipitation would not be useful, as there would not be enough time to deliver early warnings. Consequently, a flood forecast must be based on rainfall forecasts generated some hours or days before the beginning of the flood event. A statement has been included in Section 2 (Case study) of the reviewed manuscript.

6. In Table 1, please include a column that shows the sources of the data or the resolution for each input dataset (e.g., land use, precipitation, topography).

We thank the reviewer for the comment. We have extended the information in said table. Moreover, we have added Figure 3 to explain in detail the soil type and Table 2 with the associated Brooks-Corey parameters.

7. Lines 210–220 It could be a good idea to talk about how this method could be used in other Mediterranean catchments or in other areas.  

We thank the reviewer for this valuable suggestion. We agree discussing the application of this methodology to other Mediterranean or Spanish catchments is significant. In fact, we are going to extrapolate this methodology to make a flood forecasting in catchment in Aragon, Spain; which belongs to the Ebro River Basin Authority. Hence, we have added Section 5.3 to the Discussion Section.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors responded adequately to my observations. No further comments.

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents a well-designed study with clear contributions to the field and should be accepted for publication