DWD's new Seamless Integrated Forecasting system (SINFONY) is entering a new phase.

Up to now the focus was initial development until "fitness for operations" in a project framework, to improve very-short-range forecasting of intense convective events from observation time up to 12h, with weather radar at the heart of it. Having DWD's warning process (forecasters, automated systems) as well as German flood forecasting authorities in mind, seamless ensemble products in observation space (radar reflectivity composites, precipitation fields and convective cell objects) were developed, as well as informations on the probability  of hazards like heavy precipitation, hail, wind gusts and lightning.

In the last eight years we developed

1) Radar Nowcasting ensembles for areal precipitation, reflectivity (STEPS-DWD) and convective cell objects including hail and life cycle information (KONRAD3D-EPS) with good forecast quality up to 1-2 hours.

2) A new regional NWP ICON-ensemble model (ICON-RUC-EPS) with LETKF assimilation of 3D radar volumes, cell objects, Meteosat VIS and IR channels and hourly new forecasts on the km-scale, whose forecast quality exceeds Nowcasting after about 1h. Advanced model physics (2-moment microphysics with prognostic hail) lead to an improved forecast of convective clouds and to more consistent model equivalents for radar data and visible/infrared satellite data based on detailed forward operators (EMVORADO and RTTOV-MFASIS).

3) An optimal combinations ("blending") of Nowcasting and NWP ensemble forecasts in observation space, which constitute the seamless forecasts of the SINFONY. Gridded combined precipitation and reflectivity ensembles (INTENSE) are targeted towards hydrologic warnings. Combined Nowcasting- and NWP cell object ensembles (KONRAD3D-SINFONY) help evolve DWD’s warning process for convective hazards towards flexible “warn-on-objects". Reflectivity composites and 3D cell objects are computed with the same software as for observations and Nowcasting.

4) Common Nowcasting and NWP verification systems for precipitation, reflectivity and cell objects help to continuously improve the SINFONY components.

Now we will expand in two directions.

On the one hand, the existing prototypes will be installed and maintained operationally as a cross-cutting activity on a long-term basis. We already got the ICON-RUC-EPS into operations in July 2024.

On the other hand we will further improve the systems in a new project phase 2025-2028 towards seamless forecasts for other parameters/phenomena and other user groups:

• Temperature, wind, cloudiness, fog, solar radiation, visibility, ceiling for
• aeronautical forecasts, renewables, customer data portals,
• with good year-round performance,
• and seamless forecasts beyond 12h.

For this, we try to improve the ICON-RUC forecasts for these parameters, we blend ICON-RUC into ICON-D2 and we integrate satellite data into our Nowcasting and combined products. 

How to cite: Blahak, U. and the Team SINFONY & Friends: Status and perspectives of SINFONY – the seamless combination of Nowcasting and NWP at DWD, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-357, https://doi.org/10.5194/ems2025-357, 2025.

Convective events have long been one of the most difficult phenomena to predict, making them a major focus of the SINFONY project at the DWD. Our new product is at the forefront of this effort, designed to revolutionize short-range forecasting (up to 14 hours) for convective storms by seamlessly integrating enhanced nowcasting and numerical weather prediction (NWP) into one powerful, cohesive forecast tool.

In this work, we aim to combine convective cells detected through probabilistic nowcasting with those from numerical weather prediction (NWP). The detection of these cells is performed using KONRAD3D, a state-of-the-art method developed at the DWD to identify and track convective cells based on radar reflectivity. This approach can also be applied to cells simulated by NWP, as the model forward operator, EMVORADO, generates simulated radar data with the same structure and temporal resolution as actual radar observations.

First, the simulated cells are spatially clustered using the DBSCAN method. After clustering, each observed cell is linked to the nearest cluster of simulated objects. The properties of the simulated cells are then compared to those of the observed radar cells using a score known as the total interest. Only cells that exceed a certain total interest threshold—indicating the greatest similarity to the observations—are selected for combination. Finally, the selected simulated cells are spatially adjusted so that their centroid matches the position of the nearest observed cell. Simulated cells detected within a specific time window around the observation but not matched to an observed cell are excluded from further consideration.

We also perform ensemble nowcasting of observed cells. In this process, the position, movement, and severity are subjected to stochastic noise. Additionally, a parabolic lifecycle of cell severity is assumed.
As a result, each observed cell receives a seamless forecast of its development through ensemble nowcasting, as well as from assigned model cells.
Moreover, thanks to the model input, our approach can account for the formation of new cells, which offers an advantage over pure nowcasting.

For forecasters, we provide compact information in the form of representative members and occurrence probabilities for cells based on their severity.
Since last year, the product has been under evaluation not only by the DWD but also by the ESSL Testbed.
We present an overview of our product, along with statistical verification and a prominent case study.

How to cite: Josipovic, L., Strotjohann, N.-L., and Blahak, U.: Seamless Combination of Convective from Nowcasting and NWP with KONRAD3D-SINFONY – Part 1: General Aspects    , EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-565, https://doi.org/10.5194/ems2025-565, 2025.

KONRAD3D-SINFONY aims to improve the prediction of convective thunderstorms by integrating object-based nowcasting with ensemble forecasts from the high-resolution ICON-RUC model. This contribution focuses on the verification of KONRAD3D-SINFONY, beginning with a detailed evaluation of how accurately the ICON-RUC model reproduces the position, geometry, intensity, and temporal evolution of observed convective cells. In the subsequent analysis, we quantify the extent to which incorporating nowcasting data enhances these characteristics across a range of forecast lead times.

Building on the ensemble of both corrected and model-derived cells, we investigate various methods to calculate probabilistic watch regions, which mark the areas where thunderstorms are most likely to occur. In addition, we aim to visualize the most relevant cell characteristics — including intensity, size, and associated weather hazards — in a format that is both clear and easily interpretable by forecasters. The quality of this condensed forecast visualization is assessed by comparing our predictions to both detected convective cells and the operational thunderstorm warnings issued by the German Weather Service (DWD).

KONRAD3D-SINFONY is still under active development at DWD, and several physical effects and sources of information are not yet fully utilized within the current system. We highlight how incorporating these additional components could further optimize the cell combination process and improve the overall forecast quality. We conclude by outlining our plans for ongoing refinements and future development, with the goal of combining the results from nowcasting and numerical weather prediction in an optimal way and clearly presenting the most relevant information to users.

How to cite: Strotjohann, N. L., Josipovic, L., and Blahak, U.: Seamless Combination of Convective Cells from Nowcasting and NWP with KONRAD3D-SINFONY – Part 2: Verification and Visualization, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-607, https://doi.org/10.5194/ems2025-607, 2025.

MeteoSwiss is developing a unified, probabilistic, gridded forecast system designed to deliver added value for specific user groups, such as hydrological modelers. Today, these users often need to run their models for each numerical weather prediction source separately. Following the example of forecasts in the national weather app—which already delivers daily 7-day forecasts for ~6,000 locations by integrating ICON (1 and 2 km), IFS, and INCA—the new system aims to simplify this by combining all relevant information into a seamless, high-quality forecast. This enables more efficient and consistent downstream applications, improving usability, accessibility, and decision-making for both internal and external users.

A key feature is the Seamless Rapid Update Cycle (S-RUC), a data-driven model architecture that extends forecasts up to 10 days and updates short-range guidance every 10 minutes, leveraging the latest observations such as radar and satellite data to maximize predictive power.

MeteoSwiss is also developing an operational MLOps platform to standardize and accelerate the use of machine learning. This infrastructure will support model development, training, validation, and deployment, ensuring scalability and maintainability of future AI applications.

The initial focus is on temperature, precipitation, wind, and cloud cover, with an emphasis on high-impact weather and hydrological relevance for Swiss territory.

As a result, the project will lead to the consolidation of existing nowcasting and postprocessing systems. This will reduce duplicated effort, streamline visualization workflows, and lower maintenance costs and dependencies through shared tools and common data formats.

This presentation outlines MeteoSwiss’ early steps toward seamless, machine-learning-enabled forecasting and highlights the methodological and operational innovations under development.

How to cite: Moret, L. and the Seamless Weather Project Team: Toward seamless weather forecasts, MeteoSwiss’ first steps., EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-461, https://doi.org/10.5194/ems2025-461, 2025.

Within the Destination Earth initiative, the Extremes Digital Twin (DT) has been produced continuously at ECMWF since summer 2023. The Extremes DT is running at 4.4 km horizontal resolution every day starting at 00 UTC out to 5 days. Due to computational constraints it is currently comprised of a single deterministic forecast, which limits an assessment of uncertainty unless an ensemble or other methodologies are introduced. Currently, three approaches are under investigation to quantify uncertainty in the Extremes DT: (i) use of a physical ensemble at 4.4 km resolution with 10 members, (ii) use of machine learning to downscale the uncertainty information from the operational 9 km ensemble system to 4.4 km and (iii) application of statistical/neighbourhood (NB) methods for uncertainty quantification.  

In this presentation the implementation and first results of the NB method are described. The aim of the NB method is to characterize uncertainty from a single high-resolution deterministic forecast run by taking into account the inherent uncertainty that comes from location errors or timing errors of predicted localised weather phenomena. Uncertainty related to location errors is represented by constructing a pseudo-ensemble from neighbouring grid points in horizontal space. Uncertainty related to timing errors is constructed by investigating a time window incorporating forecast lead times before and after the actual lead time. Following this construction, the two main free parameters of the method that require investigation are the size of the search radius in horizontal space and the width of the time window. NB methods are widely used in kilometre-scale limited area models that typically operate on regular grids. The challenge here has been the implementation on an irregular spherical grid for global fields. It is now possible to compute NB probabilities on the native TCo2559 grid of the IFS model, so that the reduction of extreme values by interpolation is avoided.

The presentation will describe the results that aim to optimise space- and timescale choices of the neighbourhood method for applications.

How to cite: Szintai, B., Roberts, N., and Brookshaw, A.: Quantifying uncertainty from the Extremes Digital Twin of Destination Earth, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-573, https://doi.org/10.5194/ems2025-573, 2025.

The Korean Integrated Model (KIM) is the operational global Numerical Weather Prediction (NWP) system developed by the Korea Institute of Atmospheric Prediction Systems (KIAPS). Since its operational deployment in 2020, KIM has undergone numerous enhancements aimed at reducing global bias and improving forecast skill. Recently, KIAPS introduced KIM version 4.0, an 8-km horizontal resolution model, featuring several advancements such as the revised scale-aware parameters in the KSAS convection scheme. A comprehensive year-round performance evaluation demonstrated significant reductions in global errors, particularly evident in mid-latitude regions especially during the winter season. At the same time, high-impact weather predictions such as heavy rain fall, heatwave and typhoon, showed much refined prediction performance, which allows for more reliable support for forecasters during extreme weather events over the Korean Peninsula. Efforts are currently focused on preparing the next version, with an emphasis on physics modifications to further alleviate global bias. A new diagnostic cloud scheme is under evaluation to replace the existing prognostic scheme, enhancing consistency in cloud and hydrometeors simulations, which is a crucial step towards more effectively addressing persistent global biases. Additionally, adjustments in the global hydrometeor structures impacting radiation response have been made through sub-grid hydrometeors modifications in the convection parameterization scheme, alongside phase conversion updates in the WSM5 microphysics scheme. Enhanced vertical wind structures are being explored via improvements in sub-grid orography. This presentation discusses the sensitivities and major impacts of these modifications. It also introduces ongoing efforts to develop Korea’s next-generation NWP systems at KIAPS, providing insights into currents challenges and future plans. 

How to cite: Lee, E.-H., Bae, S. Y., and Koo, M.-S.: Dynamics and Physics Updates in the KIM4.0 and Future Upgrade Plans, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-122, https://doi.org/10.5194/ems2025-122, 2025.

Seasonal forecasting refers to the prediction of weather conditions over one to several months, offering insights into temperature and precipitation anomalies relative to climatological means. Positioned between short-term weather forecasting and long-term climate projections, it supports critical decision-making in sectors such as agriculture and water resource management. Despite operational advances, forecast skill remains limited in many regions, particularly for precipitation, due to challenges in model resolution, data quality and the representation of complex climate dynamics. Improving seasonal forecast systems is essential to mitigate the impacts of climate variability and support sustainable development.

The primary objective of this study is to identify the optimal combination of input datasets and cumulus convection parameterization schemes for conducting seasonal hindcast and forecast simulations using the Weather Research and Forecasting (WRF) model version 4.5. Sensitivity tests were conducted in the European domain at a horizontal resolution of 0.5° × 0.5°. A total of four simulations were carried out with changes in both the driving datasets that are used for the initial and boundary conditions, and the cumulus parameterization schemes, for the evaluation period 1998-2000. The simulations were forced by two different reanalysis products: the ECMWF Reanalysis version 5 (ERA5) and the NCEP Climate Forecast System Reanalysis (CFSR). Additionally, two cumulus parameterizations were utilized: the Grell-Freitas and the Kain-Fritsch scheme. The different simulated data were assessed to ensure the model’s robust performance. The evaluation of the simulations was performed for both temperature and total precipitation using the Climatic Research Unit (CRU) high-resolution gridded dataset.

The results showed that WRF simulates temperature quite satisfactorily in the ERA5-driven simulations, and especially in the one with Kain-Fritsch scheme, while the CFSR-driven simulations highly deviate from the gridded data. Regarding total precipitation – a rather challenging parameter that is difficult to predict in numerical weather models as it depends on many different atmospheric conditions – WRF appear to overestimate this parameter in all simulations during winter, with the overestimation being more enhanced in mountainous areas. On the contrary, during summer total precipitation is mainly underestimated in the CFSR-driven simulations, while an enhanced overestimation is observed in mountainous areas (and especially the Alps) in all four simulations. These findings suggest that using ERA5 as forcing data in conjunction with the Kain-Fritsch cumulus scheme provides a more reliable performance of WRF model for seasonal simulations in the area of study.

Acknowledgments
The work was supported by PREVENT project. This project has received funding from Horizon Europe programme under Grant Agreement No: 101081276.

How to cite: Velikou, K., Manios, E. M., Papadopoulos-Zachos, A., Tolika, K., and Anagnostopoulou, C.: Improving Seasonal Forecasts: A Sensitivity Analysis of the WRF Model Configurations, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-189, https://doi.org/10.5194/ems2025-189, 2025.

In the Sahel, the frequency and magnitude of floods caused by intense hydrometeorological events have been steadily increasing for more than three decades. These floods have significant impacts on populations and livelihoods, undermining efforts toward sustainable development. The Early Warnings for All (EW4ALL) initiative, launched by the World Meteorological Organization (WMO) and other UN agencies, identifies early warning systems as a key tool for reducing the impacts of floods and other natural disasters. One of the core components of such systems is numerical weather prediction.

In Niger and Burkina Faso, the respective National Meteorological Services (DMN and ANAM) have adopted the Weather Research and Forecasting (WRF) model within their operational forecasting chains, each using different model configurations. As part of the SLAPIS Sahel project, a capacity-building initiative was launched to support both institutions in improving the performance of their WRF-based systems for forecasting high-impact weather events.

This study presents the preliminary results of a verification process comparing the two operational WRF configurations over a computational domain covering both countries, with a horizontal resolution of 4 km. Simulations were initialized with both GFS and ECMWF global datasets and focused on the rainy seasons (July–August–September) of 2023 and 2024. Forecasts of precipitation and surface temperature were evaluated against observational data from ANAM and DMN’s official networks, satellite-based estimates (CHIRPS and TAMSAT), and ERA5-Land reanalysis.

Verification was conducted using standard point-based skill scores—such as Probability of Detection (POD), success ratio, bias, Critical Success Index (CSI), and Root Mean Square Error (RMSE)—as well as spatial metrics including the Fractional Skill Score (FSS). Moreover, significant case studies corresponding to flood events have been analyzed in detail.

The results show that high-resolution modeling allows for a more accurate spatial reconstruction of intense rainfall events compared to global-scale models. However, due to the predominantly convective nature of Sahelian rainfall, the quantitative accuracy remains insufficient for fully deterministic flood forecasting. Future work will focus on integrating the existing physically based models with post-processing techniques based on machine learning to enhance the operational prediction of extreme weather events and support early warning dissemination at local and national scales.

How to cite: Pasi, F., Bere, R. T., Adamou Sayri, Y., Tarchiani, V., and Capecchi, V.: Performance Evaluation of High-Resolution WRF Forecasts with Multi-Source Observations over the Sahel, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-307, https://doi.org/10.5194/ems2025-307, 2025.

Global lightning forecasts are of particular interest to the aviation industry. However, global weather forecasts are typically based on numerical weather prediction model (NWP) runs with resolutions that cannot resolve deep convection, hence it needs to be parameterized. Although lightning may be diagnosed in these models, e.g. with the Lightning Potential Index (LPI,1), the quality of such forecasts is only mediocre. Successful attempts have been made to produce lightning probability forecasts with neural networks that use the LPI as input features and lighting observations as target (2,3). Moreover, high resolution global lightning data is available (GLD360, Vaisala, 4) which has been used for nowcasting of thunderstorms up to 2 hours lead time (5).

This work shows how the combination of global lightning observations (GLD360 from Vaisala, 4) with NWP can significantly improve the global lightning probability forecast for up to 8 hours. A graph neural network initialized with global lightning observations of the past hours has been developed. Observations are transferred into a latent space by a fully connected neural network and then integrated forward in time. Multiple loops using a convolutional graph neural network are employed to include spatial information. Here, NWP forecasts are fed in as feature and are the driver for the initiation of new thunderstorms, their propagation and decay.

The model is trained with one year ICON Reanalysis data and applied to global ICON ensemble forecasts. Preliminary verification results based on the resolution component of the Brier Score show that initializing the model with global lightning observations of the past hours significantly outperforms a model version without that initialization. This improvement remains for lead times up to 8 hours, then the verification scores of the two model versions converge. Furthermore, case studies show that the lightning is well resolved during the first forecast hours, i.e. beyond the typical 2 hour nowcasting horizon.

References:

(1) Schröder et al., 2022: Subgrid scale Lightning Potential Index for ICON with parameterized convection. Reports on ICON (10), DOI: https://doi.org/10.5676/dwd_pub/nwv/icon_010 .

(2) Baumgartner, M., Schröder, G., and Primo, C.: Forecasting lightning probabilities derived from the Lightning Potential Index using neural networks, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-34, https://doi.org/10.5194/ems2023-34, 2023.

(3) Schröder, G., Baumgartner, M., Primo, C., and Theis, S.: Challenges in training neural networks for global lightning probability forecasts, EMS Annual Meeting 2024, Barcelona, Spain, 1–6 Sep 2024, EMS2024-475, https://doi.org/10.5194/ems2024-475, 2024.

(4) https://www.vaisala.com/en/products/systems/lightning/gld360

(5) Müller, R., Barleben, A., Haussler, S., & Jerg, M. (2022). A Novel Approach for the Global Detection and Nowcasting of Deep Convection and Thunderstorms. Remote Sensing, 14(14), 3372. https://doi.org/10.3390/rs14143372

How to cite: Schröder, G., Baumgartner, M., Primo, C., and Theis, S.: Seamless global lightning forecasts based on convolutional graph neural networks, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-62, https://doi.org/10.5194/ems2025-62, 2025.

The identification and characterization of atmospheric turbulence, particularly rotors, is crucial for aviation safety due to their small scale and turbulent nature as well as their difficulty to accurately predict. Traditional methods of detecting rotors rely on manual inspection and reports. These are limited in temporal coverage, requiring significant investment in training to ensure accuracy when observing. This study explores the opportunity to use machine learning methods to identify features within remote sensing data. In particular this study focusses on convolutional neural networks (CNNs), to identify rotors within Light Detection and Ranging (LiDAR) output. LiDAR technology provides high-resolution, three-dimensional wind field data, enabling detailed analysis of atmospheric phenomena. By leveraging this data annotated by field expertise, we developed a robust CNN model capable of detecting rotors with high accuracy. The model was trained on labeled rotor data, with a comprehensive hyperparameter search conducted to optimize its performance. The results indicate that the CNN models trained effectively, achieving high performance on the training dataset, though there was a tendency to overfit. Despite this, the ability to correctly classify rotor images, even with an overpredictive bias, remains valuable for operational meteorologists. This study demonstrates the potential of machine learning techniques to advance turbulence detection in the meteorological domain, ultimately contributing to safer aviation practices. This also opens the door for generating longer datasets that can then be combined with other data sources such as numerical weather prediction data allowing for increased understanding of atmospheric conditions conducive to their formation as well as potentially highlighting more common locations for formation, leading to better asset protection operations.

How to cite: Ramsdale, S., Ascione, I., Luo, C., and Fu, Z.: Using LiDAR Output to Identify Atmospheric Rotors: A Convolutional Neural Network Approach, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-115, https://doi.org/10.5194/ems2025-115, 2025.

Modern weather radar networks play an indispensable role in nowcasting and short-term weather forecasting. They provide high-resolution, volumetric data crucial for identifying convective structures, precipitation intensity, and storm dynamics. However, the native spherical-polar coordinate system used by radar instruments presents significant challenges when integrating this data into numerical weather prediction (NWP) models and generating standardized meteorological products. To overcome these limitations, we present a novel, modular framework for the generation of 3D Cartesian radar products derived from both single and multi-radar network data. This framework is designed for use in operational meteorology, data assimilation, and research applications.

The system transforms raw radar observations into a unified, geospatially referenced Cartesian grid, enabling the production of key volumetric weather products such as Constant Altitude Plan Position Indicator (CAPPI), Vertically Integrated Liquid (VIL), and Echo Tops. Developed both as a command line service and standalone Windows Forms application using C# , the framework is equipped with configurable modules for radar data acquisition, spatial transformation through interpolation, product visualization, and standardized data export. It supports both proprietary and open radar data formats and offers flexible configuration of domain size, resolution, and compositing strategies.

In operational testing, the system generated 72 distinct product types, demonstrating consistent performance and scalability. Benchmarking revealed that processing times scale linearly with domain volume and the number of radar sources, which confirms the suitability of the approach for near-real-time deployment. Additional capabilities include automated data acquisition, dynamic product scheduling, and improved interpolation schemes that enhance spatial fidelity.

By bridging the gap between raw radar data and application-ready volumetric products, the framework significantly improves radar data accessibility and usability. It facilitates better situational awareness for forecasters and supports integration into NWP workflows. Its modular, extensible design lays the foundation for future developments, including full automation and real-time nowcasting integration across national and regional radar networks.

How to cite: Szuster, P. and Kołodziej, J.: Framework for generation of 3D weather radar data composite products, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-162, https://doi.org/10.5194/ems2025-162, 2025.

IMPROVER (Integrated Model Post-Processing and Verification) has been developed by the Met Office as an open-source probability-based post-processing system to fully exploit our convection permitting, hourly cycling ensemble forecasts. Post-processed MOGREPS-UK model forecasts are blended with deterministic UKV model forecasts and data from the coarser resolution global ensemble, MOGREPS-G as well as ECMWF, to produce seamless probabilistic forecasts from now out to 14 days. For precipitation, an extrapolation nowcast is also blended in at the start. Forecasts are converted to probabilities at the start, and all initial stages of post-processing are performed on gridded data, with site-specific forecasts extracted as a final step, helping to ensure consistency. Data are processed on a 10km global grid and on a 2km UK-centred grid. Physical and statistical corrections are applied to the data to ensure the probability distribution functions for each source model are sufficiently similar for blending into a seamless probabilistic forecast.

In this talk we present an overview of the enhancements added to the Met Office IMPROVER capability over the past year, including applications of statistical and machine learning methods and comparison with truth data to improve parameters including replacing Ensemble Model Output Statistics (EMOS) with Statistical Anomaly Model Output Statistics (SAMOS) for temperature, wind speed and wind gust. We also consider Quantile Regression Random Forests visibility, extending forecasts to 14 days using ECMWF data and working towards being able to create physically consistent realisations of precipitation parameters for hydrological models, including regridding surface snow amount to reflect a higher-resolution grid and orography.

How to cite: Moseley, S., Ayliffe, B., Evans, G., Hooper, B., Hurst, K., and White, M.: Recent updates to the Met Office IMPROVER post-processing system, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-60, https://doi.org/10.5194/ems2025-60, 2025.

India received a lot of rain from the two main rain-bearing systems (i.e., Monsoon Depressions (MDs) and Deep Depressions (DDs)) during the monsoon season. The MDs and DDs contributed roughly 60% of the seasonal rainfall and primarily originated over the Bay of Bengal (BoB) before moving northwest to the mainland. As they migrate from the BoB to the mainland, they have significantly impacted the eastern coast of India, especially Odisha. Due to its heavy reliance on intricate, non-linear processes at a smaller scale, rainfall forecasting is very challenging to predict with a reasonable degree of accuracy, particularly at the district level.  For the first time, we attempted to employ high-resolution real-time forecast outputs (i.e., WRF) as the input feature to district-scale DL models to reduce prediction error rather than relying on observation or reanalysis datasets.

 In this study, we used input data from the Weather Research and Forecast (WRF) forecast at high resolution (3km), which was initialized using GFS real-time forecast, to enhance the spatial and categorical rainfall prediction (intensity) at the district scale by introducing two DL-based architectures:  U-Net (+A) (attention-based U-Net architecture) and KU-Net (+A) (Attention-based Kernelized U-Net Architecture). The model was trained using 24 cases of monsoon LPS (12 MDs and 12 DDs) spanning between 2007 and 2018 and tested for two cases, July 2023 and August 2023, over Odisha in real-time.

Our suggested model (KU-Net (+A)) significantly enhanced the Odisha district-level rainfall prediction. While WRF shows the MAE of more than 25 mm and 36 mm for both instances, respectively, the DL models decreased the MAE values to less than 8 mm up to Day 4 for case 1 and less than 15 mm for case 2. The rainfall distribution for case 1 demonstrates that the suggested DL model better captures the spatial rainfall, whereas the WRF underestimates it by less than 10 mm. In the second scenario, WRF underestimates the HREs, while the suggested DL model accurately predicts them in terms of size, intensity, and dispersion. The confusion matrix indicates that there were no HREs in Case 1, and the DL model's TPR (True Positive Rate) for the LR (Light Rain) and MR (Moderate Rain) classes is greater (>80%) than WRF's (52.5%). Only the HR (Heavy Rain), VHR (Very Heavy Rain), and HER (Extremely Heavy Rain) categories' TPRs of 85.7%, 87%, and 100%, respectively, are captured by the suggested DL model in the second real-time scenario. The study's conclusions directly impact how early warning systems can improve using the DL model.

How to cite: Trivedi, D., Pattnaik, S., and Sharma, O.: Reducing the real-time heavy rainfall forecast error associated with Monsoon depressions and deep depressions over the Indian region using deep learning, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-262, https://doi.org/10.5194/ems2025-262, 2025.

National weather services play a crucial role in mitigating the impact of severe weather events by issuing timely warnings for conditions such as frost (prolonged sub-threshold temperatures) or heavy rain (intense rain over a defined period). Automating the production of weather warnings has the potential to improve the consistency and uniformity of warnings while improving operational efficiency. Here, we present advancements in a prototype designed to automate critical processes within the new warning system of the German Meteorological Service (Deutscher Wetterdienst, DWD), which is developed in the program RainBoW ("Risk-based, Application-oriented and INdividualizaBle Provision of Optimized Warning Information"). This work introduces an algorithm designed to derive consistent weather events from forecast members drawn from various model versions and types.

By standardizing event detection within ensemble forecasts, the prototype aims to improve the accuracy and coherence of automated warnings. The system currently integrates real-time meteorological data from the numerical weather model ICON in four setup variations (rapid update cycle and normal cycle of local area version, EU version and global version). It applies rule-based decision-making to assess severe weather conditions within individual ensemble members and aggregates the various outputs into a unified result. The objective is to generate seamless warning information with lead times of up to seven days. By processing ensemble members from different model variants collectively while allowing adjustable weighting of inputs, we ensure a balanced and adaptable mechanism that accounts for variations in model performance. Preliminary results suggest that improvements can be expected especially in complex terrain, where the characteristics of deep valleys and high peaks are much better captured when deriving frost warnings. Therefore, this enhancement has the potential to increase user benefit by delivering more precise and actionable warnings.

How to cite: Königer, L., Felsberg, A., Hammelmann, J., Schröder, G., Baumgartner, M., Klink, M., and Feige, K.: Deriving Event-Based Warnings from Ensemble Forecasts, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-371, https://doi.org/10.5194/ems2025-371, 2025.

The impact and mitigation of the impact of weather on the public and prosperity of the UK is a core focus of the Met Office. At the heart of this is the National Severe Weather Warnings Service (NSWWS), the UK wide impact-based warnings service used in planning and preparation by responders as well as the wider public. Traditionally this process requires significant human effort from expert meteorologists and consultation with stakeholders to understand potential impacts from severe weather. With increasing volumes and performance of numerical weather prediction (NWP) data available to operational meteorologists a new approach is required to mitigate the cognitive load increasingly put on their shoulder. With an impact based warnings system this allows us to move from weather parameter forecasting to recognition of when whether may be impactful as a starting point. This identification can then be used to direct further analysis, limiting the requirement for an operational meteorologist to look at all data to make their initial assessment and instead direct them to relevant hazard analysis. To explore this new paradigm this work looks at creating a prediction of whether NSWWS warnings are likely to be issued from coarse resolution NWP parameters alone, mimicking the starting point of human processes, limiting the data volumes necessary for this initial assessment but allowing increased exploitation of ensemble information. The models developed show skill in identifying the potential of NSWWS warnings on a regional scale though performance by region and parameter does vary. This suggests that this approach has merit but this is a supporting model for directing human exploration, supporting the value of the operational meteorologist in years to come.

How to cite: Ramsdale, S.: A Machine Learned approach to the Likelihood of UK Impact Based Severe Weather Warnings from Coarse Resolution Synoptic Weather Regimes, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-119, https://doi.org/10.5194/ems2025-119, 2025.

The German Meteorological Service (Deutscher Wetterdienst, DWD) is renewing its weather warning system through a dedicated program called RainBoW ("Risk-based, Application-oriented and INdividualizaBle Provision of Optimized Warning Information"). With the overarching goal to more strongly support informed decision-making in the face of significant weather situations, we are specifically working to improve the comprehensibility of warnings to make them more effective, and to extend their forecast horizon to give recipients more time to act. Additionally, targeted at users with specialized warning requirements, we aim to enable the individualization of warnings. 

RainBoW's outcomes will be released gradually. The first increment focuses on the goal to improve the comprehensibility of warnings for the general public, which includes the introduction of a consistent warning criteria catalogue. DWD's warnings currently operate on a four-level scale, but the amount of available levels varies between warned weather elements. For example, there are four warning levels for wind gusts, while only three are used for rain. Moreover, the current least severe level does not always represent hazardous weather, which might lead to unnecessary alarmism. To address these discrepancies, we plan to introduce a three-level system with a consistent semantic definition for each of the levels. A second aspect to advance the comprehensibility of warnings is a restructured warning text with a focus on potential weather impacts and recommended actions, which we also plan to release within the first increment. 

Besides this, RainBoW's first increment will address the goal of extending the forecast horizon of warnings, initially focussing on thunderstorm warnings. Currently, warnings are issued once a thunderstorm is detected by nowcasting systems, resulting in very short lead times. In case of expected severe thunderstorms, a pre-warning is provided earlier, but this is not the case for all available warning levels. To bridge this gap, we plan to introduce a new thunderstorm warning, which uses the forecasted thunderstorm potential as a base instead of the immediate detected thunderstorm in nowcasting. These new thunderstorm warnings are intended to provide earlier information before a short-notice warning is issued, which is at the same time more reliable than the pre-warning.   

While the first increment focusses on the goals to improve the comprehensibility of warnings and to extend their forecast horizon, we are also working on RainBoW's upcoming increments. Amongst other improvements, this comprises advancing the individualization of warnings via the so-called warning portal ("DWD-Warnportal"), and the implementation of an automatic warning trend covering up to seven days. 

This contribution will go into detail concerning the contents of the first increment, and will give a brief overview of other ongoing work within RainBoW. 

How to cite: Feige, K. and the RainBoW Team: RainBoW's first increment: Renewing the weather warning system at the German Meteorological Service one step at a time, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-443, https://doi.org/10.5194/ems2025-443, 2025.

The DWD warning portal (“DWD-Warnportal”) is a web app developed by the German Meteorological Service (Deutscher Wetterdienst, DWD), targeting users with specialized warning information needs related to their particular weather-dependent applications. The portal thus complements the standardized DWD weather warnings, which are based on a catalogue of fixed hazard-related warning criteria. The DWD warning portal user group, for instance from the field of disaster management, needs warning information beyond fixed warning thresholds. The goal to provide individualizable warning information is embedded in the program RainBoW (“Risk-based, Application-oriented and INdividualizaBle Provision of Optimized Warning Information”), which is renewing the warning system of the DWD. 

In this contribution, we present the current state of the DWD warning portal prototype. The web app provides warning information for a set of preconfigured warning thresholds, which may differ from the standardized DWD weather warnings. Besides this, users can customize the evaluated area, weather model, and spatial aggregation measures. In addition to the individualization aspect, another key advantage of the warning portal is that the derivation of weather warning information is based on ensemble data. Hence, all provided results include the respective occurrence probabilities, allowing the users to make more informed decisions.

To obtain a final product tailored to the needs of the target group, the DWD warning portal team follows a cyclic development strategy. First, a lightweight prototype is implemented. Then, access to this prototype is granted to a growing group of test users, representing a broad set of targeted application areas. In annual workshops and follow-up questionnaires, feedback is gathered from these test users, which is then used as a guidance to prioritize which features should be implemented next. Then, the cycle starts again. After implementing a DWD warning portal configurator, which will enable users to save their own persistent warning profiles, plans for future releases include providing push options of warning reports on various media channels

How to cite: Riedl, K., Vogel, C., Rüth, M., Reetz, B., Schaab, R., Noel, L., Niebuhr, H., and Feige, K.: Individualized Weather Information in the DWD Warning Portal, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-389, https://doi.org/10.5194/ems2025-389, 2025.

The automatization of weather warnings is an ongoing endeavor and plays an important role in the program RainBoW ("Risk-based, Application-oriented and INdividualizaBle Provision of Optimized Warning Information"), which aims to renew the warning system of the German Meteorological Service (Deutscher Wetterdienst, DWD). Automated warnings are often generated at relatively high frequency, depending on the availability of updated forecast data. For a fixed forecast date, these updates of the warning information may lead to frequent changes of, for example, the warning level, that could confuse end-users. One method to circumvent frequent jumps in the warning level is called smoothing. This work presents a prototype system for temporal smoothing of automated weather warnings, focusing on wind gusts and thunderstorm events to gain temporal stability and to minimize fluctuating warnings.

For wind gusts, this study uses automatically generated warnings based on ensemble wind gust forecasts from the in-house ICON model, which is updated hourly. Smoothing of thunderstorm warnings is based on the output of the NowCastMIX nowcasting application, which predicts the development of thunderstorm cells and their characteristics in 5-minute increments for up to one hour. The techniques to enhance the temporal warning stability are applied to a grid-based framework. For thunderstorm and wind gust warnings, temporal smoothing is achieved by evaluating the maximum warning level across a series of the most recent forecast data for the same forecast date. Wind-related parameters that complement the warning, including gust speeds and wind direction, undergo a weighted smoothing procedure that accounts for the higher precision of more recent forecasts. This methodology aims to reduce unnecessary fluctuations in warning outputs without compromising the responsiveness of the system to genuine changes in weather conditions.

This prototype for smoothing wind gust warnings contributes to the ongoing effort to automatize weather warnings at the DWD and is an important step towards more user-friendly warning-products.

How to cite: Hammelmann, J., Brune, S., Felsberg, A., Königer, L., Schröder, G., Baumgartner, M., Klink, M., and Feige, K.: Temporal Smoothing of Automated Wind Gust and Thunderstorm Warnings, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-196, https://doi.org/10.5194/ems2025-196, 2025.

How to cite: Bhend, J., Zeman, C., Knirsch, L., and Mahlstein, I.: Insights from the Objective Verification of Weather Warnings , EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-246, https://doi.org/10.5194/ems2025-246, 2025.

We present initial results from the INLINE project funded by the Union Civil Protection Mechanism (UCPM). The objective is to develop probabilistic short-term warning products of severe weather with focus on localized phenomena, such as heavy rainfall, lightning and wind gusts in sub-hourly and kilometer-scale. The products are implemented in two different configurations: pan-European and one adapted to Finland by using locally available data sources.

The forecast models developed in INLINE are based on deep learning and extrapolation techniques. Our primary tool is the Simpler yet Better Video Prediction (SimVP): a deep learning model originally developed for prediction of multi-channel digital image sequences (Tan et al. 2025). The SimVP predictions suffer from progressive loss of small-scale features and extreme values with increasing lead time. Thus, we propose a modification to address this shortcoming. As an alternative we consider the extrapolation models implemented in pySTEPS (https://pysteps.github.io) and compare their forecast skill and computational performance with SimVP. We conclude that SimVP has up to 20% better forecast skill for lead times in the 2-4 hour range. This is mainly because of its ability to predict growth and decay of weather phenomena and spatiotemporal correlations between multiple variables. For very short time ranges (under 30 minutes), extrapolation methods still remain as a viable alternative to deep learning. In addition, we apply the perturbation generators implemented in pySTEPS to produce realistic forecast ensembles in a computationally efficient manner.

Radar-derived rain rate from the EUMETNET OPERA radar network is used as the main forecast variable. Additional variables include convective available potential energy (CAPE), convective inhibition (CIN) and wind u- and v-components. For training the models, we use ERA5 reanalyses, and use the corresponding IFS forecasts from ECMWF for real-time prediction. In the configuration localized to Finland, lightning flash density gridded to 10 km spatial resolution and 10-minute time window is used as an additional variable.

Gridded forecasts are translated into multi-hazard warnings by using combined thresholds that are defined by user-specified logical operations (e.g. rain rate over 10 mm/h and wind speed over 25 m/s). We present initial evaluation of the warnings by case studies from major storm events during 2023 and 2024. This is done using the European Severe Weather Database (ESWD) reports and emergency calls available from the Finnish PRONTO database as independent verification observations. Our results highlight the challenges in pan-European prediction of weather hazards, such as regional variability of data quality and availability of verification observations.

How to cite: Pulkkinen, S. and Myllykoski, H.: Pan-European severe weather warnings by using deep learning and extrapolation techniques, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-317, https://doi.org/10.5194/ems2025-317, 2025.

Machine learning offers a vast range of applications, including weather and hazard forecasting. The ability of these methods to more easily and efficiently extract information from diverse and novel data types enables the transition ty. This study demonstrates the feasibility of this transition using an operational forecasting system. Data on human and natural ignitions were integrated along with observed fire activity. This enabled the data-driven models to reduce the persistent overprediction of fire danger in fuel-limited biomes. This resulted in fewer false alarms and more informative outputs compared with traditional methods.

A key factor driving this improvement has been the availability of global datasets for fuel dynamics and fire detection These datasets were not accessible during the development of earlier physics-based models. Three models with increasing complexity (random forest, XGBoost and neural networks) were used in a set of ablation experiments to evaluate the importance of data compared to the complexity of machine learning (ML) architecture , progressively incorporating additional data sources during model training. Combining all data sources yields the best fire activity predictions, both globally and regionally. From this ideal scenario, prediction skill degrades by roughly 30% when using only weather or ignition data and 15% with only fuel data (for the XGboost). Similar decreases are obtained also with the other ML architectures. Fuel data is especially important, as it captures the effects of weather on vegetation. Using any two out of the three data sources improves prediction quality, reducing the degradation to between 17% and 13% relative to only using one source.

We found that the enhanced predictive skill of ML models stems largely from the comprehensive characterization of fire processes provided by these datasets, rather than from the complexity of the ML methods themselves. Our findings highlight the critical importance of high-quality training data in improving forecast accuracy. While the rapid advancement of ML techniques generates good and feasible results, there is a risk of undervaluing the essential role of data acquisition and, where necessary, its creation through physical modeling. Our results underscore that investing in robust datasets is indispensable and should not be overlooked in the pursuit of complex algorithms.

How to cite: Wetterhall, F., Di Giuseppe, F., McNorton, J., and Lombardi, A.: Global data-driven prediction of fire activity requires good quality data, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-687, https://doi.org/10.5194/ems2025-687, 2025.

Floods are one of the most common natural disasters, and the rate of flood-related disasters has more than doubled since the year 2000. Consequently, accurate and timely warnings are critical for mitigating flood risks, especially for major events impacting thousands. While recent advancements in global hydrological models offer significant potential for widespread discharge predictions, their operational use in warning systems is often limited by the difficulty of validating forecasts in individual ungauged locations.

To address this limitation, we present a global early warning system for major flood events that builds on top of discharge predictions from the Global Hydrological Model (Nearing et al., 2024). Our core concept shifts the focus from predicting and validating the hydrological predictions to the targeted identification of severe events.

Our approach employs a two-step algorithm. First, an agglomerative clustering algorithm is used to group discharge predictions that exceed predefined return period thresholds, resulting in extended spatiotemporal clusters representing potential flood events. Second, a supervised machine learning classification model is trained to identify clusters with a high likelihood of major flooding based on their properties, such as the affected area, estimated affected population, and severity at individual gauges. For training and validation, we utilize the Dartmouth Flood Observatory (DFO) and GDACS datasets, providing records of historical flood events.

Leveraging this approach, we were able to enhance our global early warning capabilities for severe flood events, increase coverage and extend reach to previously uncovered regions. Our results underscore the potential of our approach to improve global early warning for severe flood events.

How to cite: Ikan, T., Coehn, D., and Gilon, O.: SEA - Global Early Warning System for Severe Flood Events, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-154, https://doi.org/10.5194/ems2025-154, 2025.

Visualization is an important and ubiquitous tool in the daily work of weather forecasters and atmospheric researchers to analyse data from simulations and observations. At the EMS 2024, we presented the state of our ongoing efforts to release a new version 2.0 of our domain-specific meteorological visualization tool Met.3D (documentation including installation instructions available at https://met3d.readthedocs.org).

Met.3D is an open-source effort to make interactive, 3-D, feature-based, and ensemble visualization techniques accessible to the meteorological community. Since the public release of version 1.0 in 2015, Met.3D has been used in multiple visualization research projects targeted at atmospheric science applications, and has evolved into a feature-rich visual analysis tool facilitating rapid exploration of gridded atmospheric data. The software is based on the concept of “building a bridge” between “traditional” 2-D visual analysis techniques and interactive 3-D techniques powered by modern graphics hardware. It allows users to analyse data using combinations of feature-based displays (e.g., atmospheric fronts and jet streams), “traditional” 2-D maps and cross-sections, meteorological diagrams, ensemble displays, and 3-D visualization including direct volume rendering, isosurfaces and trajectories, all combined in an interactive 3-D context.

This year, we present the new version 2.0 of Met.3D. The tool has become more stable (major code revisions and bug fixes), usable (featuring a new user interface, a drag-and-drop-based interactive workflow, and a batch-production mode), and integrates new visualization techniques not commonly available in other visualization tools (including interactive feature-based visualization of jet-streams and fronts, as well as interactive trajectory-based flow visualization). In this presentation, we present an overview of the updates to the software and show how it can be freely used by the community for research, forecasting, and teaching tasks.

How to cite: Rautenhaus, M., Fischer, C., Vogt, T., Beckert, A., and Hartz, M.: A new Met.3D version 2.0: Rapid exploration of gridded atmospheric data with interactive 3-D visualization, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-369, https://doi.org/10.5194/ems2025-369, 2025.