BSBora Shehu

Bora Shehu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Bora Shehu has a total of 23 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Can LSTM Neural Hydrology help us trade space for time in rainfall-induced megafloods?

    by Philipp Tanzeglock, Bora Shehu

    The catastrophic July 2021 flood in the Ahr Valley (Rhineland-Palatinate, Germany) highlights the urgent need to improve our understanding and modelling of rainfall-induced megafloods. Conventional conceptual hydrological models often fail to accurately simulate flood peaks during such extreme events. Owing to their very rare occurrence, megafloods are typically absent from calibration periods, as available discharge observations are too short in time. In contrast, the spatial coverage of discharge observations is steadily increasing. Hydrologically and physiographically similar catchments may therefore provide valuable information on flood response behavior that has not yet been observed in the catchment of interest. In this study, we investigate whether spatial information can compensate for limited temporal observations by applying a long short-term memory (LSTM) neural network within the Neural Hydrology framework (Kratzert et al., 2021), which is capable of learning patterns from large datasets and transferring them to similar, yet distinct, hydrological settings.For this purpose, in this study, we use a large dataset of catchments across Central Europe and Germany with observed discharge and meteorological data from 1970 onwards to model hourly discharge at Ahr Valley. A series of experiments is designed using different combinations of temporal coverage and sets of physiographically similar catchments to evaluate their ability to reproduce flood behavior at Ahr Valley. The methodological framework consists of two steps: (i) training the Neural Hydrology model on a set of similar catchments (excluding the Ahr catchment) using split-sample validation, and (ii) validating the trained models for extreme flood events, including the 2021 megaflood, at several Ahr sub-catchments.By systematically comparing different configurations of spatial and temporal information, we address the following questions: Can time be successfully traded for space when simulating the 2021 Ahr megaflood? How can hydrologically similar catchments be identified most effectively? And can neural hydrology outperform the conventional conceptual models used operationally for the Ahr event (LARSIM and HBV-Light)

  • Open Access

    Breaking the Limits: Stress Testing Hydrological Models Beyond Observed Extremes

    by Bora Shehu, Philipp Tanzeglock, Patricio Yeste, Paul Voit, Maik Heistermann, Axel Bronstert

    How far can hydrological models be pushed before they break down? This study explores that question by exposing a range of modelling approaches—simple conceptual models such as the Direct Runoff Model, more complex conceptual models like HBV-Light and LARSIM, as well as a data-driven model based on LSTM networks —to both real and hypothetical rainfall extremes. Beyond reproducing extreme rainfall events that caused historical floods (Ahr, Münster, and Elbe), the models are subjected to deliberately exaggerated and synthetic rainfall scenarios that challenge the physical and conceptual limits of model design. These “stress runs” reveal how each model responds when rainfall becomes exceptionally intense, prolonged, or short but extreme — conditions that are increasingly relevant under a changing climate.As a case study, results for the Ahr catchment at an hourly resolution are presented, including analyses of different initial states and calibration periods. By examining model robustness—the ability to produce physically plausible runoff across a wide spectrum of extreme conditions—and identifying failure modes, the study uncovers hidden sensitivities, structural biases, and nonlinear behaviors that standard validation approaches may overlook. The goal is to rethink how robustness is assessed and to advance hydrological models capable of withstanding the extremes of the future.

  • Open Access

    Challenges in analysing and modelling extreme floods: The case study of Ahr catchment

    by Bora Shehu, Axel Bronstert

    The extreme weather conditions of July 2021 caused major flooding’s in multiple tributaries of the Meuse and Rhine rivers. Particularly the Ahr Valley in Germany was greatly affected, where exceptional damage and severe human loss was registered. Since then, several studies have been conducted to understand the extremity, the major driving forces, the particular mechanisms of this flood and the possible impacts of climate change on the generation of such an event. Here the main objective is to perform a hydrological analysis of the July 2021 Ahr event and discuss the challenges in modelling or analysing this event.First, we show that particularly the rainfall field is associated with high uncertainty, as seen by the high variability between the different rainfall products available. The average areal rainfall volume can differ between products with as much as 50mm/day, which constitutes almost 55% of the rainfall volume estimated by Radolan. To capture the full uncertainty-range of the rainfall field, rainfall simulations conditioned both on radar and station observations are implemented. Next, based on rainfall simulations and reconstructed discharge data, runoff coefficients (Rc) are shown to be ranging between 0.6 to 1.2 (median 0.7). These values are clearly higher than expected in continental climate (Rc ~ 0.20-0.51) and the latest 100-year return flood observed in 2016 (with Rc ~ 0.4). The high lower range suggests, that the dominant processes have changed, with slower components of surface runoff shifting to faster ones. This agrees well with the observed traces of erosion, surface water and flow paths in parts of the catchment.Lastly, the reconstructed discharge data are also subjected to uncertainty due to lack of observations and the non-representativeness of the stage-discharge curve during the flood wave. Hence, high Rc values do not only originate from underestimated rainfall but as well from possible overestimated flood volume. For this purpose, discharge was estimated with the Larsim model. As expected due to the change of the dominant processes, the pre-event parameter set underestimates considerably the flood volume, while the post-calibration one agrees better with the reconstructed data. On both cases, the computed runoff coefficient ranges between 0.4 to 0.7.To conclude, extraordinary events such as the July 2021 in Ahr Catchment, are accompanied with high uncertainty and as such are difficult to be analysed or modelled. Nevertheless, the results dictate that the surface runoff played an important role in July 2021. At the same time, it is clear that the landscape still has a considerable retention effect, between 30% and 50%, even for such a heavy rainfall.

  • Open Access

    Flood frequency analysis using mean daily flows vs. instantaneous peak flows

    by Anne Bartens, Bora Shehu, Uwe Haberlandt

    In many cases, flood frequency analysis (FFA) needs to be carried out on mean daily flows (MDF) instead of instantaneous peak flows (IPF), which can lead to underestimation of design flows. Typically, correction methods are applied to the MDF data to account for such underestimation. In this study, we first analyse the error distribution of MDF-derived flood quantiles over 648 catchments in Germany. The results show that using MDF instead of IPF data can lead to underestimation of the mean annual peak flow (MHQ) by up to 80 % and mainly depends on the catchment area but appears to be influenced by gauge elevation as well. This relationship is shown to differ for summer vs. winter floods. To correct such underestimation, different linear models based on predictors derived from MDF hydrograph and catchment characteristics are investigated. Apart from the catchment area, a key predictor in these models is the event-based ratio of flood peak to flood volume (p/V ratio) obtained by the MDF data. The p/V models applied to either MDF-derived events or statistics seem to outperform other reference correction methods. Moreover, they require a minimum data input, are easily applied, and are valid for the entire study area. The best results are achieved when the L moments of the MDF maximum annual series are corrected with the proposed model, which reduces the flood quantile errors by up to 60 %. The approach behaves particularly well in smaller catchments (<500 km2), where reference methods fall short. However, the limit of the proposed approach is reached for catchment sizes under 100 km2, where the hydrograph information from the daily series is no longer capable of approximating instantaneous flood dynamics and gauge elevations below 100 m, where the difference between MDF and IPF floods is very small.


  • Open Access

    Uncertainty estimation of regionalised depth–duration–frequency curves in Germany

    by Bora Shehu, Uwe Haberlandt

    The estimation of rainfall depth–duration–frequency (DDF) curves is necessary for the design of several water systems and protection works. These curves are typically estimated from observed locations, but due to different sources of uncertainties, the risk may be underestimated. Therefore, it becomes crucial to quantify the uncertainty ranges of such curves. For this purpose, the propagation of different uncertainty sources in the regionalisation of the DDF curves for Germany is investigated. Annual extremes are extracted at each location for different durations (from 5 min up to 7 d), and local extreme value analysis is performed according to Koutsoyiannis et al. (1998). Following this analysis, five parameters are obtained for each station, from which four are interpolated using external drift kriging, while one is kept constant over the whole region. Finally, quantiles are derived for each location, duration and given return period. Through a non-parametric bootstrap and geostatistical spatial simulations, the uncertainty is estimated in terms of precision (width of 95 % confidence interval) and accuracy (expected error) for three different components of the regionalisation: (i) local estimation of parameters, (ii) variogram estimation and (iii) spatial estimation of parameters. First, two methods were tested for their suitability in generating multiple equiprobable spatial simulations: sequential Gaussian simulations (SGSs) and simulated annealing (SA) simulations. Between the two, SGS proved to be more accurate and was chosen for the uncertainty estimation from spatial simulations. Next, 100 realisations were run at each component of the regionalisation procedure to investigate their impact on the final regionalisation of parameters and DDF curves, and later combined simulations were performed to propagate the uncertainty from the main components to the final DDF curves. It was found that spatial estimation is the major uncertainty component in the chosen regionalisation procedure, followed by the local estimation of rainfall extremes. In particular, the variogram uncertainty had very little effect on the overall estimation of DDF curves. We conclude that the best way to estimate the total uncertainty consisted of a combination between local resampling and spatial simulations, which resulted in more precise estimation at long observation locations and a decline in precision at unobserved locations according to the distance and density of the observations in the vicinity. Through this combination, the total uncertainty was simulated by 10 000 runs in Germany, and it indicated that, depending on the location and duration level, tolerance ranges from ± 10 %–30 % for low-return periods (lower than 10 years) and from ± 15 %–60 % for high-return periods (higher than 10 years) should be expected, with the very short durations (5 min) being more uncertain than long durations.


  • Open Access

    Comparing uncertainty propagation of different methods for regionalized IDF curves in Germany

    by Bora Shehu, Uwe Haberlandt

    Design extreme rainfall maps are essential for the construction of many water systems and works, and are typically achieved by regionalizing extreme rainfall statistics from ground-based observations. Different methods are used for such task, where the most popular are kriging and index-based regionalization. In a previous study conducted in Germany, Shehu et al. (2022) revealed that kriging with external drift performed better than index-based regionalization in terms of accuracy (smaller error obtained from cross-validation), however it is still unclear which of the method is superior in terms of precision (wideness of prediction intervals). As the risk may be underestimated due to different sources of uncertainty, a more certain method (in terms of narrower prediction intervals) is preferable (while maintaining a good accuracy). Therefore, the objective of this study is to investigate the propagation of different uncertainty sources for both kriging and index-based regionalization and compare these two in terms of precision and accuracy. To conduct this study, around 1200 ground-based observations at fine temporal scales (5min) from the German Weather Service (DWD) for whole Germany are employed. For each ground-based observation the annual maximum volumes at different durations (from 5mins up to 7days) are extracted, and local IDF curves are estimated according to Koutsoyiannis et al. (1998). For spatial uncertainty evaluation in the kriging system sequential Gaussian simulation (sGs) together will local sample bootstrapping are employed as shown in Shehu and Haberlandt (2022). On the other hand, the uncertainty in index-based regionalization is evaluated based on a combination of regional sample bootstrapping and spatial simulations of the index. The precision of IDF curves from both methods in terms of 95% confidence interval width is compared on a cross-validation procedure at the locations with more than 40 years of observation. The results of this study reveal how the uncertainty of annual rainfall extremes propagates from local estimation to the regionalization of IDF curves based on kriging and index-based regionalization. The comparison of the uncertainty in terms of precision sheds light on which method can produce narrower prediction intervals and hence is more precise in regionalizing IDF curves. Additionally, the accuracy of both methods is advised in order to discuss the advantages and disadvantages of each method for generating spatial IDF curves.References: Koutsoyiannis, D., Kozonis, D. and Manetas, A.: A mathematical framework for studying rainfall intensity-duration-frequency relationships, J. Hydrol., 206(1–2), 118–135, doi:10.1016/S0022-1694(98)00097-3, 1998.Shehu, B., Willems, W., Stockel, H., Thiele, L., and Haberlandt, U.: Regionalisation of Rainfall Depth-Duration-Frequency curves in Germany, Hydrol. Earth Syst. Sci. Discuss. [preprint], https://doi.org/10.5194/hess-2022-118, in review, 2022.Shehu, B. and Haberlandt, U.: Uncertainty estimation of regionalised depth–duration–frequency curves in Germany, Hydrol. Earth Syst. Sci. Discuss. [preprint], https://doi.org/10.5194/hess-2022-254, in review, 2022.

  • Open Access

    Regionalisation of rainfall depth–duration–frequency curves with different data types in Germany

    by Bora Shehu, Winfried Willems, Henrike Stockel, Luisa-Bianca Thiele, Uwe Haberlandt

    Rainfall depth–duration–frequency (DDF) curves are required for the design of several water systems and protection works. For the reliable estimation of such curves, long and dense observation networks are necessary, which in practice is seldom the case. Usually observations with different accuracy, temporal resolution and density are present. In this study, we investigate the integration of different observation datasets under different methods for the local and regional estimation of DDF curves in Germany. For this purpose, two competitive DDF procedures for local estimation (Koutsoyiannis et al., 1998; Fischer and Schumann, 2018) and two for regional estimation (kriging theory vs. index based) are implemented and compared. Available station data from the German Weather Service (DWD) for Germany are employed, which includes 5000 daily stations with more than 10 years available, 1261 high-resolution (1 min) recordings with an observation period between 10 and 20 years, and finally 133 high-resolution (1 min) recordings with 60–70 years of observations. The performance of the selected approaches is evaluated by cross-validation, where the local DDFs from the long sub-hourly time series are considered the true reference. The results reveal that the best approach for the estimation of the DDF curves in Germany is by first deriving the local extreme value statistics based on Koutsoyiannis et al.'s (1998) framework and later using the kriging regionalisation of long sub-hourly time series with the short sub-hourly time series acting as an external drift. The integration of the daily stations proved to be useful only for DDF values of a low return period (T[a] < 10 years) but does not introduce any improvement for higher return periods (T[a] ≥ 10 years).


  • Open Access

    Comparison of rainfall generators with regionalisation for the estimation of rainfall erosivity at ungauged sites

    by Ross Pidoto, Nejc Bezak, Hannes Müller-Thomy, Bora Shehu, Ana Claudia Callau-Beyer, Katarina Zabret, Uwe Haberlandt

    Rainfall erosivity values are required for soil erosion prediction. To calculate the mean annual rainfall erosivity (R), long-term high-resolution observed rainfall data are required, which are often not available. To overcome the issue of limited data availability in space and time, four methods were employed and evaluated: direct regionalisation of R, regionalisation of 5 min rainfall, disaggregation of daily rainfall into 5 min time steps, and a regionalised stochastic rainfall model. The impact of station density is considered for each of the methods. The study is carried out using 159 recording and 150 non-recording (daily) rainfall stations in and around the federal state of Lower Saxony, Germany. In addition, the minimum record length necessary to adequately estimate R was investigated. Results show that the direct regionalisation of mean annual erosivity is best in terms of both relative bias and relative root mean square error (RMSE), followed by the regionalisation of the 5 min rainfall data, which yields better results than the rainfall generation models, namely an alternating renewal model (ARM) and a multiplicative cascade model. However, a key advantage of using regionalised rainfall models is the ability to generate time series that can be used for the estimation of the erosive event characteristics. This is not possible if regionalising only R. Using the stochastic ARM, it was assessed that more than 60 years of data are needed in most cases to reach a stable estimate of annual rainfall erosivity. Moreover, the temporal resolution of measuring devices was found to have a significant effect on R, with coarser temporal resolution leading to a higher relative bias.


  • Open Access

    Uncertainty estimation of regionalised depth–duration–frequency curves in Germany

    by Bora Shehu, Uwe Haberlandt

    The estimation of rainfall depth-duration-frequency (DDF) curves is necessary for the design of several water systems and protection works. These curves are typically estimated from observed locations, but due to different sources of uncertainties, the risk may be underestimated. Therefore, it becomes crucial to quantify the uncertainty ranges of such curves. For this purpose, the propagation of different uncertainty sources in the regionalisation of the DDF curves for Germany is investigated. Annual extremes are extracted at each location for different durations (from 5mins up to 7days), and local extreme value analysis is performed according to Koutsoyiannis et al. (1998). Following this analysis, five parameters are obtained for each station, from which four are interpolated using external drift kriging, while one is kept constant over the whole region. Finally, quantiles are derived for each location, duration and given return period. Through a non-parametric bootstrap and geostatistical spatial simulations, the uncertainty is estimated in terms of precision (width of 95 % confidence interval) and accuracy (expected error) for three different components of the regionalisation: i) local estimation of parameters, ii) variogram estimation and iii) spatial estimation of parameters. First two methods were tested for their suitability in generating multiple equiprobable spatial simulations: sequential Gaussian simulations (SGS) and simulated annealing (SA) simulations. Between the two, SGS proved to be more accurate and was chosen for the uncertainty estimation from spatial simulations. Next, 100 realisations were run at each component of the regionalisation procedure to investigate their impact on the final regionalisation of parameters and DDFs curves, and later combined simulations were performed to propagate the uncertainty from the main components to the final DDFs curves. It was found that spatial estimation is the major uncertainty component in the chosen regionalisation procedure, followed by the local estimation of rainfall extremes. In particular, the variogram uncertainty had very little effect in the overall estimation of DDFs curves. We conclude that the best way to estimate the total uncertainty consisted of a combination between local resampling and spatial simulations, which resulted in more precise estimation at long observation locations, and a decline in precision at un-observed locations according to the distance and density of the observations in the vicinity. Through this combination, the total uncertainty was simulated by 10,000 runs in Germany, and indicated, that depending on the location and duration level, tolerance ranges from ±10–30 % for low return periods (lower than 10 years), and from ±15–60 % for high return periods (higher than 10 years) should be expected, with the very short durations (5min) being more uncertain than long durations.


  • Open Access

    Regionalisation of Rainfall Depth-Duration-Frequency curves in Germany

    by Bora Shehu, Winfried Willems, Henrike Stockel, Luisa Thiele, Uwe Haberlandt

    Rainfall depth-duration-frequency (DDF) curves are required for the design of several water systems and protection works. These curves are typically generated from the station data by fitting a theoretical distribution to the annual extremes (AMS). The aim of this study is to investigate the use of different data types and methods for estimating reliable DDF curves covering whole Germany. The following three questions are investigated for the evaluation and regionalisation of the DDF curves in Germany: i) which is the best local estimation method, ii) which regionalisation method shows best performance, and iii) which data sets should be used and how they should be integrated. For this purpose, two competitive DDF-procedures for local estimation (Koutsoyiannis et al. 1998, Fischer and Schumann, 2018) and two for regional estimation (kriging theory vs index-based) are implemented and compared. Available station data from the German Weather Service (DWD) for Germany are employed, which includes; 5000 daily stations with more than 40 years available, 1261 high resolution (1 min) recordings with observations period between 10 and 20 years, and finally 133 high resolution (1min) recordings with 60–70 years of observations. The performance of the selected approaches is evaluated by cross-validation, where the local DDFs from the long sub-hourly time series are considered the true reference. The results reveal that the best approach for the estimation of the DDF curves in Germany is by first deriving the local extreme value statistics based on Koutsoyiannis et al. 1998 framework, and later use the kriging regionalisation of long sub-hourly time series with the short sub-hourly time series acting as an external drift. The integration of the daily stations proved to be useful only for DDF values of very low return period (T<10 years), but not doesn’t introduce any improvement for higher return periods (T≥10 years).