Gauge-independent, multi-source precipitation merging methods are well-established approach for improving precipitation estimates. These methods predominantly aim to minimise uncertainty in precipitation magnitude, yet they frequently neglect errors in distinguishing between rain and no-rain events. This oversight often leads to biased merging weights and suboptimal precipitation estimates. In this study, we introduce an innovative two-stage framework called the Generalised Signal-to-Noise Ratio Optimisation (G-SNR) framework, specifically designed to address these limitations. The first stage employs the Categorical Triple Collocation-Merging (CTC-M) method for binary merging, effectively mitigating errors in rain/no-rain classification. The second stage applies Signal-to-Noise Ratio Optimisation (SNR-opt) to enhance precipitation magnitude estimates, leveraging the improved classification outcomes. Evaluation results demonstrate that G-SNR consistently surpasses both input data and existing methods in terms of binary classification and magnitude estimation. Importantly, it achieves error reductions across all percentiles, delivering robust performance for both low and extreme precipitation events. This framework provides a comprehensive and reliable solution to longstanding challenges in precipitation merging, significantly enhancing both accuracy and dependability.

How to cite: Kim, S., Shah, S., Liu, Y., and Sharma, A.: Improving Precipitation Merging: A Generalized Two-Stage Framework Using the Signal-to-Noise Ratio Optimization (SNR-opt), EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-530, https://doi.org/10.5194/egusphere-egu25-530, 2025.

Excess runoff from heavy precipitation events (HPEs) in urban environments often leads to urban flooding, a severe hazard with significant implications for human life, property, and infrastructure. Modeling runoff response in complex and heterogeneous urban areas, while accounting for rainstorm and surface characteristics, remains a significant challenge. Climate change and urbanization are key drivers of increased future urban runoff intensity. Research on the interaction between these factors and urban runoff in the eastern Mediterranean region is particularly limited. Previous studies using high-resolution models have projected an increase in short-duration rainfall intensities, alongside a decrease in long-duration intensities, rainfall coverage area, and total event rainfall during HPEs in the eastern Mediterranean under the RCP8.5 scenario. The current study examines the implications of these changes on peak discharge and volume of urban runoff by the end of the 21st century and evaluates the influence of varying urbanization scenarios, providing insights into the interplay between climate change and urban development. Using high-resolution radar-rainfall and surface data, we developed and calibrated a SWMM-based urban rainfall-runoff model for the Nahal Ra'anana basin (13 km²) on Israel's coastal plane. This Mediterranean-climate region encompasses most of the city of Ra'anana and has approximately 40% impervious surfaces. The model was developed using 23 runoff events utilizing leave-one-out cross-validation and a multi-objective optimization approach, and demonstrated robust performance with KGE values of 0.80 for runoff peak discharge and 0.83 for total runoff volume. A variance-based sensitivity analysis identified three primary factors influencing urban runoff: rainstorm intensity distribution, impervious surface coverage, and basin water storage. Analysis of HPEs under historical and future climatic conditions revealed that, at the current urbanization level of the city, climate change alone is unlikely to alter peak or total runoff discharge significantly. This is attributed to the decrease in total event rainfall and coverage area, alongside an increase in short-duration rainfall intensities. However, with substantial urbanization (e.g., increasing impervious surface to 52% or more), future climate HPEs are expected to exhibit a noticeable shift in the trend, leading to increased peak discharge. Further analysis indicates the elevated importance of rainfall intensities in determining runoff peaks in future climate conditions. In historical HPEs the maximum rainfall intensities over a 60-minute duration strongly correlate with peak runoff discharge (R²=0.75), where in future climate HPEs, correlations of shorter and longer rainfall durations are improved compared to historical HPEs with the maximum obtained for 60–120-minute durations (R²=0.81). The non-linear discharge response to climate change underscores the importance of integrating climate projections into urban planning to mitigate future flooding risks and highlight the potential for short-term peak discharge forecasting under both current and future climatic conditions.

How to cite: Nussbaum, R., Armon, M., and Morin, E.: Urban runoff response to climate-change-driven heavy precipitation and urbanization, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-841, https://doi.org/10.5194/egusphere-egu25-841, 2025.

This study explores the feasibility of using the information contained in observed streamflow discharge measurements to inversely correct catchment-average precipitation time series provided by reanalysis products. We explore this possibility by training LSTM models to predict precipitation. The first model uses discharge as an input feature along with other meteorological factors, while the second model uses only the meteorological factors. Although the model provided with discharge information showed better mean performance, a detailed analysis of various time series measures across the continental scale revealed underestimation biases when compared with the original reanalysis product used for training. However, an out-of-sample test showed that the inversely estimated precipitation is better able to reproduce small-scale, high-impact events that are poorly represented in the original reanalysis product. Further, using the inversely generated precipitation time series for classical hydrological “forward” modeling resulted in improved estimates for streamflow and soil moisture. Given the notable disconnect between reanalysis products and extreme events, particularly in data-scarce regions worldwide, our findings have implications for achieving better estimates of precipitation associated with high-impact events.

How to cite: Manoj J, A., Loritz, R., Gupta, H., and Zehe, E.: Can discharge be used to inversely correct precipitation?, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4086, https://doi.org/10.5194/egusphere-egu25-4086, 2025.

To properly adapt to climate change, we need to estimate extreme precipitation probability in future climate scenarios. The task is particularly challenging for sub-daily and sub-hourly extremes, as they are hardly represented by most of the available climate models. As an alternative to explicit model simulations, one can use stochastic models trained on physical covariates. For example, it was recently shown that we can predict changes in sub-daily and sub-hourly extreme precipitation only based on shifts in wet-day daily temperatures. With the aim of extending the applicability of such stochastic models, we examine here the use of covariates representing both thermodynamic and dynamic processes.

We focus on a set of ~300 stations in the Alps (from France, Switzerland, Austria, Italy) for which we have sub-daily precipitation and temperature observations. First, we assess the importance of statistical independence of the events on the identification of the scaling relationships between extreme precipitation and temperature that are commonly used to quantify the thermodynamic component. Then, we evaluate the relative importance of the thermodynamic and dynamic components for durations ranging between 10 minutes and 24 hours using as covariates dew point, vertical velocity at 500 hPa, and divergence at 300 hPa from ERA5 reanalysis simulations.

Our results show that (1) evaluating extreme precipitation-temperature scaling relations using all the wet time intervals (as done in several studies) leads to biased estimates of the scaling rates relevant for extreme sub-daily precipitation projections. (2) The scaling rates between extreme precipitation and dew point tend to decrease logarithmically with duration, an information that can be used to extract the scaling rate at sub-hourly durations from hourly observations. (3) The importance of the thermodynamic component decreases with duration (rank correlation decreases from ~0.55 at 10 minutes to ~0.2 at 24 hours), while the importance of the dynamic component that can be appreciated at the ERA5 resolution (~30 km) tends to increase with duration (rank correlation increases from ~0.2 at 10 minutes to ~0.45 at 24 hours). (4) From a stochastic simulation perspective, temperatures and dew point during precipitation events in the Alps can be simulated using generalized normal distributions (or normal distributions in case of seasonal data), while vertical velocities and divergence need to be simulated using skewed models such as a generalized extreme value distribution. 

How to cite: Marra, F., Ciceri, R., Stante, S., and Sada, C.: Toward the stochastic modelling of extreme precipitation probability with thermodynamic and dynamic covariates, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4866, https://doi.org/10.5194/egusphere-egu25-4866, 2025.

  This study proposes a method utilizing Rainfall Transition Probability (RTP) to create rainfall scenarios that consider the temporal distribution of heavy rainstorm events. RTP refers to the probability of rainfall amount at time t occurring after a specific rainfall amount at time t+1. The method consists of a temporal distribution that builds region-specific RTPs using rainfall data observed at 1-minute interval, a function that users define the desired conditions for the rainfall scenario, and a processing module that generates scenarios based on the RTP. To develop the RTP, the rainfall data about 1-minute interval used for separating Independent Rainstorm Events (IREs) according to each region. Among the identified IREs, those exhibiting high-intensity rainfall (above 15 mm/hour) are used to calculate and establish the RTP. Afterward, users define the conditions for the rainfall scenario in the function with conditions such as region, total rainfall, and rainfall duration. The generator then utilizes the RTP for the selected region to generate various rainfall scenarios with different temporal distributions and presents them to the user. By extracting the temporal distribution from regional IREs, the generator reflects local rainfall patterns and can be applied to regional hydrological modelling.

Keywords: Rainfall Generator, Rainfall Transition Probability, 1-minute Rainfall Data, Temporal Distribution, Heavy Rainstorm Events

Acknowledgment

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00334564).

How to cite: Cha, H., Baik, J., Lee, J., Na, W., and Jun, C.: Development of Rainfall Scenario with Transition Probability Reflecting on Temporal Distribution of Heavy Rainstorm Events, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5084, https://doi.org/10.5194/egusphere-egu25-5084, 2025.

Abstract

In this study was investigated three different microphysics schemes on the rainfall patterns over Kuwait on 02 January 2022. The primary objective is to improve precipitation predictions using the Weather Research and Forecasting (WRF) high resolution 4 km model, which has been dynamically downscaled from the Community Climate Model version 4 (CCM4). The performance of three selected microphysics schemes—Lin, WSM6, and Thompson was evaluated using the ERA5 reanalysis dataset. ERA5 has been previously validated in this region and has consistently provided reliable results, making it a suitable dataset for such studies. Three numerical simulations were conducted using the WRF model, each incorporating one of the three microphysics schemes. The simulations were assessed by comparing the model outputs against the ERA5 data to determine the accuracy of the rainfall forecasts. The results revealed that the WRF Single-Moment 6-class microphysics scheme (WSM6) outperformed the other microphysics schemes, including Lin and Thompson, in forecasting rainfall patterns during the storm. The Lin scheme was found to be the least reliable, providing less accurate rainfall predictions compared to the Thompson and WSM6 schemes. This study highlights the critical role of selecting appropriate microphysics schemes for accurate precipitation prediction, particularly in extreme weather events like the 2022 storm in Kuwait. The findings suggest that the WSM6 scheme is a more effective choice for rainfall forecasting in this region, whereas the Lin scheme may not be as suitable for this particular type of storm event. Further research is recommended to extend this analysis to different regions and storms for more comprehensive results.

How to cite: Alsarraf, H.: Evaluation of WRF Microphysics Schemes for Precipitation Forecasting in an Arid Region: A Case Study Over Kuwait, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5123, https://doi.org/10.5194/egusphere-egu25-5123, 2025.

The availability of accurate long-term gap-free precipitation data at high spatiotemporal resolutions is crucial for hydroclimatic extremes assessment, water resources management, infrastructure design, hydrological modeling, and evaluation of climate change impacts. However, many ground precipitation data contain gaps, which can hinder accurate assessments and analyses. Therefore, different gridded precipitation products (PPs) are promising alternatives to overcome this deficiency, especially in heterogeneous regions with different terrains where ground observations are sparse or non-existent. This study evaluates four daily precipitation products, i.e., SPARTACUS, IMERG-V07, CHIRPS-V2.0, and ERA5-land, to determine their performance in representing observed patterns, the intensity, and frequency of extreme precipitation events in Kamp Catchment in Austria for the period 1998-2020 at different temporal scales. The Kamp River is the longest in the “Waldviertel” region and has key ecological, societal, and economic functions, with many popular leisure and excursion destinations for tourism. The catchment also frequently experiences severe floods, causing adverse socioeconomic impacts. Ground-based precipitation data from 33 stations distributed within and around the catchment are used to conduct point-to-pixel evaluation for the four products. To measure the disparity between the products and the ground observations, six performance metrics were used: the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Bias Error (MBE), Nash-Sutcliffe Efficiency (NSE), Correlation coefficient (r), and Willmott index of agreement (d). The intensity and frequency of extreme precipitation reflected by the four evaluated PPs are assessed using selected extreme climate indices defined by the Expert Team on Climate Change Detection and Indices (ETCCDI). The PPs were ranked to select the best-performing product in the study area. The ranking results of the performance metrics revealed that SPARTACUS is the top-performing product on a daily and monthly scale and in capturing the frequency and intensity of precipitation extremes, followed by IMERG-V07 and ERA5-land, whereas CHIRPS-V2.0 ranked the lowest. SPARTACUS showed superior performance across the catchment with the highest correlation with the observed data and lowest bias (on daily and monthly scales, mean r values are 0.92 and 0.96 and mean MBE values are -0.02 and -0.81, respectively). Other products exhibit systematic precipitation underestimation. Regarding capturing precipitation extremes, all products show low skills and overestimate the daily extreme precipitation events, with the highest NSE of -0.32 shown in SPARTACUS. CHIRPS-V2.0 and ERA5-land presented different performances for detecting the longest wet and dry spells in the catchment. CHIRPS-V2.0 overestimated the consecutive dry days (CDD) and underestimated the consecutive wet days (CWD), whereas ERA5-land shows the opposite pattern. SPARTACUS shows minor overestimation of CDD and underestimation of CWD (MBE = -0.09 and 0.13 mm, respectively). Accordingly, a simple drought assessment was performed in the catchment using SPARTACUS data and the Standardized Precipitation Index (SPI). Our results highlight the importance of site-specific validation before using any precipitation products.

This study is conducted within the frame of the DISTENDER project (EU Horizon-ID 101056836), where climate extremes and climate change impacts upon several European catchments are analyzed and robust adaptation strategies are developed.

Keywords: Precipitation extremes, Precipitation products, Point-to-pixel evaluation, SPI, Kamp catchment, Austria

How to cite: Babker, Z., Zagar, M., G. Reichenau, T., Basheer, M., and Schneider, K.: Comparison and evaluation of different precipitation products in capturing climate extremes in Kamp Catchment, Austria, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6812, https://doi.org/10.5194/egusphere-egu25-6812, 2025.

Precipitation downscaling, particularly at convection-permitting scales (less than 4 km), is highly uncertain. This is especially pronounced in mountainous regions due to the interplay of complex topography and atmospheric dynamics. It impedes reliable estimation of variability and risks in localized extreme rainstorms. Deep learning-based downscaling methods have gained increasing attention but have primarily focused on deterministic prediction, which fails to capture uncertainty. Here we developed a novel Probabilistic High-resolution Precipitation Downscaling Network (P-HRDNet) with prior knowledge of key precipitation characteristics to design its loss function and model architecture. This knowledge includes data imbalance, skewed distribution, heteroscedasticity, and spatial and temporal dependencies of precipitation. P-HRDNet was tested in the southeastern Tibetan Plateau, a mountainous region lacking high-resolution precipitation data. Ten-year WRF simulations with nested domains provided coarse (9 km) and fine resolution (1 km) daily precipitation to train P-HRDNet. Compared with a baseline model SRCNN, P-HRDNet achieved greater accuracy in terms of root mean square error, mean absolute error, and Pearson correlation coefficient. Besides, it offers better uncertainty coverage and narrower uncertainty widths. This superiority is particularly evident in the extreme values. Our study highlights the importance of incorporating prior knowledge of precipitation characteristics into deep learning, and has a potential to physically constrain Artitifical-Intelligience (AI) based weather forecasting models. Furthermore, our WRF-AI framework offers an efficient solution for obtaining reliable high-resolution precipitation estimates in poorly gauged regions.

How to cite: Li, D., Yang, L., Pan, B., Liu, Y., and Zhou, Y.: Prior knowledge-constrained deep learning for probabilistic precipitation downscaling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7599, https://doi.org/10.5194/egusphere-egu25-7599, 2025.

Unlike regional climate models, convection-permitting models (CPMs) are able to resolve convection-scale processes and therefore better estimate short-duration, extreme precipitation events, particularly useful for the urban drainage community. Despite their state-of-the-art capabilities, bias correction of CPMs is still required to ensure their output is representative of the station scale, a resolution needed by many urban drainage models. Due to its simplicity, quantile-mapping is commonly used for bias-correction and downscaling, but does come with limitations that have not yet been evaluated for CPMs. This study tests five variations of empirical quantile-mapping to bias-correct and downscale the 2.2 km simulations of COSMO-CLM (a CPM) for over 70 weather stations in Switzerland. Ten years of simulation data are corrected using ten years of observations at the 30-minute interval. Traditional QM and several advanced versions are evaluated, including: using a 91-day moving window to account for temporal variability, spatial pooling of surrounding grid cells for spatial variability, and extending the observational record (to 30 years) for data variability. These techniques are validated using cross-validation and through evaluation of historical rainfall indices (e.g., consecutive dry days) and the climate change signal. Findings show that wet biases in raw CPM output remain (up to 30-35 mm/hour above the 98^th quantile) and only the moving window technique (and its combination with spatial pooling) is able to reduce biases in quantiles above the 98^th. All QM methods do reduce remaining biases, but can distort the climate change signal, particularly in indices related to frequency of rainfall. Despite the additional computational burden, the moving window technique is highly recommended to the urban drainage community as a robust technique for CPM downscaling. As more CPM simulations become available, future work will reexamine these aspects for a range of CPMs, time periods, and simulation domains.

How to cite: Cook, L., Nguyen, T., Dietzel, A., and Velasquez, P.: Correction of Precipitation Bias from Convection-Permitting Models at the Station Scale in Switzerland, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8629, https://doi.org/10.5194/egusphere-egu25-8629, 2025.

Understanding changes in global oceanic precipitation remains challenging due to limitations in current observational datasets and model deficiencies, particularly in the representation of cloud and precipitation properties within oceanic regions. To address this, we examined climatologies and trends in oceanic precipitation between 2001 and 2020 using a collection of 27 state-of-the-art satellite and reanalysis datasets available on a uniform daily 1°×1° resolution from the Frequent Rainfall Observations on Grids (FROGS) database. The results showed that reanalysis datasets generally report higher annual mean daily precipitation than satellite datasets. The tropical region exhibits the greatest absolute discrepancies in precipitation rates, while arid regions such as the southeast Pacific and Atlantic show significant relative differences among products. An increasing trend is primarily observed in satellite products, whereas reanalyses suggest strong declines. Taken together, reanalyses show pronounced decreases over the Intertropical Convergence Zone (ITCZ) and North Atlantic, contradicting the “wet gets wetter, dry gets drier” (WWDD) pattern. In contrast, the satellites better align with the WWDD pattern, with over half of oceanic regions meeting this expectation. The precipitation trend in the combined reanalysis products also exhibits the weakest consistency with sea surface temperature (SST) trends in wet regions (34.2%), compared with dry regions in the reanalysis cluster (53.4%) and both wet (59.6%) and dry (58.5%) regions in the satellite cluster. We recommend using an ensemble of satellite products for investigating global oceanic precipitation while exercising greater caution when utilizing reanalysis datasets.

How to cite: Cheng, S., Alexander, L., and Sherwood, S.: Decadal climatology and trends in global oceanic precipitation from 27 satellite and reanalysis datasets, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9341, https://doi.org/10.5194/egusphere-egu25-9341, 2025.

Despite significant advancements in the Poisson cluster-based Bartlett-Lewis model for effectively reproducing rainfall extremes, there is still room for further refinement. This study proposes a refined model, referred to as RBL7, introducing module k with a modified equation for rainfall disaggregation. This adjustment allows the power of the sine function to vary inversely with rain cell duration, thereby capturing the realistic characteristics of rainfall extremes, which often come with high intensity over short durations. Furthermore, an improved calibration approach is also proposed for the first module of the RBL7 model. This involves a hybrid optimization technique combining Particle Swarm Optimization (PSO) and fmincon methods, iterately executed until the objective function reaches the pre-assigned threshold. While the calibration of the RBL7 model relies solely on observed rainfall aggregated at hourly and longer timescales, it effectively reproduces rainfall extremes from uncalibrated sub-hourly to supra-hourly aggregation intervals, outperforming existing models using sine-2 and rectangular pulse shapes. Additionally, this refined model maintains its capability to capture rainfall standard statistics, i.e., mean, variance, covariance, skewness, and proportion of wet period, at various timescales ranging from 5 minutes to a month. These findings highlight the robustness of the RBL7 model in simulating rainfall characteristics, especially extreme values at sub-hourly aggregation intervals.

Acknowledgement

This study was supported by Korea Environment Industry & Technology Institute (KEITI) through R&D Program for Innovative Flood Protection Technologies against Climate Crisis Program (or Project), funded by Korea Ministry of Environment(MOE)(RS-2023-00218873).

How to cite: Vuong Tai, C., Park, J., Wang, L.-P., and Kim, D.: Bartlett-Lewis based stochastic rainfall model: An improvement to effectively reproduce sub-hourly rainfall extremes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14679, https://doi.org/10.5194/egusphere-egu25-14679, 2025.

Climate change is an essential part of sustainable development challenges in developing countries. Climate change represents one of the greatest environmental, social, and economic threats facing the world today. Accurate meteorological and hydrological projections are vital for effective climate adaptation and resource management, particularly under changing climate scenarios. However, the coarse spatial resolution of General Circulation Models (GCMs) limits their applicability for localized impact assessments. This study proposes a deep learning-based super-resolution approach combined with an advanced hydrological model to downscale and enhance the spatial resolution of three GCM datasets—GFDL-CM4, GISS-E2-1-G, and IPSL-CM6A-LR—to approximately 0.01°. The performance of the method is evaluated based on mean square error (RMSE), mean absolute error (MAE), Peak signal-to-noise ratio (PSNR), and Pearson correlation coefficient (R). This study hypothesizes to have more precise and accurate meteorological and hydrological predictions and projections under this framework. The model is conducted on historical climate data and compared with high-resolution observational datasets, showcasing its ability to capture fine-scale climatic and hydrological variability. This approach bridges the resolution gap in climate projections and provides a robust framework for better-informed decision-making in climate change adaptation and mitigation strategies.

Funding

This research was supported by Disaster-Safety Platform Technology Development Program of the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT. (No. 2022M3D7A1090338).

How to cite: Huong, O. S. and Lee, G.: Improving Climate Change  Data through Deep Learning Super-Resolution Downscaling of GCMs for Precise Hydrological Projections, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15576, https://doi.org/10.5194/egusphere-egu25-15576, 2025.

Precipitation time series are used as input for hydrological modeling. As the main driver of the hydrological cycle, they directly influence soil moisture, runoff, river flows, and groundwater recharge. High-resolution precipitation data is required to obtain accurate hydrological models. In addition, data should be available from different locations to reflect spatial dependencies in these models. As precipitation is measured only at selected locations, the simulated series can be used for design purposes.

In recent years, various models have been developed based on the Fourier Transform because of its ability to preserve desirable statistical properties. The concept is to transform the time series from the time domain to the frequency domain and calculate the two main components of the transformed series: the power spectrum (the square of the absolute values of the Fourier frequencies) and the phase spectrum (phase angle of the frequencies). The main idea behind all the Fourier-based models is to preserve the power spectrum because it relates to the autocorrelation function and overall structure.

This study compares the most common Fourier-based time series generators using different measures. As most spectral methods are iterative, this can be challenging for the precipitation time series, especially for the hourly resolution. In this regard, a non-iterative method is introduced. This method takes advantage of the Wiener–Khinchin theorem for the transformation between the autocorrelation function and the power spectrum. Another method, the Phase Annealing method, is introduced for precipitation time series generation and keeping the spatial and temporal properties. The results have been compared for the developed models and the most common Fourier-based methods.

How to cite: Mehrvand, M. and Bárdossy, A.: Comparative study of spectral methods for precipitation time series generators based on the conserving observed spatial and temporal properties, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16085, https://doi.org/10.5194/egusphere-egu25-16085, 2025.

With the increasing frequency and intensity of extreme rainfall events, the importance of nowcasting to minimize damage from disasters such as flash floods is becoming ever more prominent. However, most nowcasting models use loss functions aimed at minimizing the average prediction error. As a result, they tend to underestimate extreme rainfall—which has relatively low occurrence frequency but significant impact. In this study, we applied various types of weighted loss functions to a ConvLSTM-based nowcasting model to more accurately predict extreme rainfall. In particular, we varied parameters within these weighted loss functions and conducted sensitivity analyses to identify the optimal weighting strategies. We also categorized extreme rainfall types and evaluated the models’ predictive performance with weighted loss functions, thereby examining both the accuracy and stability of the model’s forecasts under extreme conditions from multiple perspectives. The results showed that the model employing a weighted loss function achieved significantly improved accuracy in predicting extreme rainfall, compared to an unweighted model. Furthermore, depending on the type of weighted loss function and parameter settings, the model demonstrated notably strong performance for specific types of extreme rainfall. This finding suggests that, in a rainfall environment characterized by high variability, dynamically selecting weighted loss functions according to forecasting objectives and conditions can enhance both the efficiency and reliability of extreme rainfall prediction. The approach presented in this study can be applied to flood forecasting and is anticipated to contribute to the advancement of deep learning–based disaster response systems, reducing the potential damage caused by natural disasters.

Acknowledgements

This work was supported by Korea Environment Industry & Technology Institute(KEITI) through R&D Program for Innovative Flood Protection Technologies against Climate Crisis Program(or Project), funded by Korea Ministry of Environment(MOE)(RS-2023-00218873).

How to cite: Choi, H., Kim, Y., and Kim, D.: Enhancing Extreme Rainfall Nowcasting with Weighted Loss Functions in Deep Learning Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19416, https://doi.org/10.5194/egusphere-egu25-19416, 2025.

Regional climate models (RCMs) are crucial tools for understanding and predicting climate change and its impacts, such as precipitation extremes. We investigate the characteristics of hourly precipitation and the associated extremes in RCM ensembles with two resolutions: km-scale (the CORDEX-FPS Convection ensemble with ~3 km grid spacing, where deep convection is represented explicitly), and coarser-scale (~12 km grid spacing, with parameterized convection). The km-scale ensemble is downscaled from the coarser one, and both cover three time periods: evaluation, historical, and end-of-the-century period under the RCP8.5 warming scenario (2000-2009, 1996-2005, and 2090-2099, respectively). Evaluating the model ensembles against data from 179 weather stations in Austria, we study how the intensity, duration, and the time of onset of precipitation depend on mean daily temperature. We then examine how these characteristics change under warming conditions.

It is well established that over the Alps the coarser RCMs produce too much light and persistent precipitation which is triggered too early in the day. We find that these shortcomings in models with parameterized convection become more pronounced with rising temperatures. We show that the km-scale ensemble closely matches observations and greatly outperforms the coarser ensemble in capturing the investigated hourly precipitation characteristics, especially at higher temperatures and on days with heavy rainfall. As high temperatures are expected to become more common in future climates, our results imply that coarser RCMs suffer from more severe biases in hourly precipitation in the future than under present climate conditions, especially for short-duration extremes. 

In this light, we also assess the ability of both km-scale and coarser RCM ensembles to capture the Clausius-Clapeyron scaling of extreme precipitation with temperature, and discuss how model deficiencies in the coarser ensemble affect this relationship.

In summary, our findings highlight the importance of km-scale RCMs for accurate simulations of hourly precipitation and its extremes, particularly in the warming climate.

How to cite: Medvedova, A., Kohlhauser, I., Maraun, D., Rotach, M. W., and Ban, N.: Hourly Precipitation Biases and Clausius-Clapeyron Scaling in Convection-Resolving and Convection-Parameterizing Regional Climate Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15818, https://doi.org/10.5194/egusphere-egu25-15818, 2025.