Sebastian Lerch
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Sebastian Lerch has a total of 134 co-authors. Here are the top co-authors by number of shared publications:
- Jieyu Chen
Institute of Statistics, Nanjing University of Information Science and Technology, Nanjing, China
14 shared publications Nina Horat
10 shared publications- Annette Möller
Faculty of Business Administration and Economics Bielefeld University Bielefeld Germany
9 shared publications - Peter Knippertz
Karlsruhe Institute of Technology (KIT), Institute of Meteorology and Climate Research Troposphere Research (IMKTRO), Karlsruhe, Germany
9 shared publications - Sándor Baran
Faculty of Informatics University of Debrecen Debrecen Hungary
9 shared publications - Benedikt Schulz
Karlsruhe Institute of Technology, Karlsruhe, Germany
7 shared publications
Latest Publications
- Open Access
- 5 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-25
Comment on egusphere-2026-4514
Data sets in the geosciences continue to grow in size, and data compression methods can facilitate efficient data storage, transfer, and utilization. Related workflows in geoscience modelling typically involve data generation (e.g. via weather forecasting), followed by compression, transfer, decompression, extraction of local subsets, and integrations into downstream tasks such as hydrological modelling. Recently, data-driven models have been introduced into geoscientific workflows with great success, allowing input data sources and types to be used in much more flexible ways than in classical numerical models, e.g. by directly using the compressed data. In this study, we therefore address the questions on 1) how various lossy compression algorithms compare in terms of general properties such as computational effort and their ability to extract spatial subsets directly from the compressed representation, and 2) how they compare in terms of compression efficiency and reproduction error. We focus on dimension reduction methods, as the resulting low-dimensional representations require fewer input channels in downstream data-driven models, thereby reducing computational costs and the risk of overfitting. We compare several dimension reduction methods in an application to five years of hourly air temperature and rainfall data, in the form of 2-d gridded spatial fields spanning a 100,000 km² domain in central Germany. The methods we compare are: Block Averaging, Principal Component Analysis, an Autoencoder, and the Ramer-Douglas-Peucker algorithm. We measure compression by the number of distinguishable objects in the compressed representation, rather than by file size, keeping in mind the potential use of the compressed data as input to data-driven models. This approach directly relates dimension reduction/compression to the number of input channels of a data-driven model, which is typically limited to avoid overfitting. Our results indicate that the methods differ substantially in terms of training and computational overhead, preservation of field-scale statistics, and the possibility to extract spatial subsets directly from the compressed representation. All methods demonstrated very good compression efficiency, with low reconstruction errors even for 99 % compression. For the spatially smooth temperature fields, Principal Component Analysis performed best, followed by the Autoencoder. For the spatially heterogeneous rainfall fields, the Autoencoder and the Ramer-Douglas-Peucker algorithm were most effective. The choice of the best method therefore depends on the specific application. This study provides a step towards a more efficient use of geoscientific data in downstream applications, either by directly using compressed data as input to data-driven models, or by efficiently extracting spatial subsets from compressed data, and using these subsets as input to local numerical models.
- 5 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-25
Open AccessComment on egusphere-2026-4514
Data sets in the geosciences continue to grow in size, and data compression methods can facilitate efficient data storage, transfer, and utilization. Related workflows in geoscience modelling typically involve data generation (e.g. via weather forecasting), followed by compression, transfer, decompression, extraction of local subsets, and integrations into downstream tasks such as hydrological modelling. Recently, data-driven models have been introduced into geoscientific workflows with great success, allowing input data sources and types to be used in much more flexible ways than in classical numerical models, e.g. by directly using the compressed data. In this study, we therefore address the questions on 1) how various lossy compression algorithms compare in terms of general properties such as computational effort and their ability to extract spatial subsets directly from the compressed representation, and 2) how they compare in terms of compression efficiency and reproduction error. We focus on dimension reduction methods, as the resulting low-dimensional representations require fewer input channels in downstream data-driven models, thereby reducing computational costs and the risk of overfitting. We compare several dimension reduction methods in an application to five years of hourly air temperature and rainfall data, in the form of 2-d gridded spatial fields spanning a 100,000 km² domain in central Germany. The methods we compare are: Block Averaging, Principal Component Analysis, an Autoencoder, and the Ramer-Douglas-Peucker algorithm. We measure compression by the number of distinguishable objects in the compressed representation, rather than by file size, keeping in mind the potential use of the compressed data as input to data-driven models. This approach directly relates dimension reduction/compression to the number of input channels of a data-driven model, which is typically limited to avoid overfitting. Our results indicate that the methods differ substantially in terms of training and computational overhead, preservation of field-scale statistics, and the possibility to extract spatial subsets directly from the compressed representation. All methods demonstrated very good compression efficiency, with low reconstruction errors even for 99 % compression. For the spatially smooth temperature fields, Principal Component Analysis performed best, followed by the Autoencoder. For the spatially heterogeneous rainfall fields, the Autoencoder and the Ramer-Douglas-Peucker algorithm were most effective. The choice of the best method therefore depends on the specific application. This study provides a step towards a more efficient use of geoscientific data in downstream applications, either by directly using compressed data as input to data-driven models, or by efficiently extracting spatial subsets from compressed data, and using these subsets as input to local numerical models.
- Open Access
- 3 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-25
Comment on egusphere-2026-4514
Data sets in the geosciences continue to grow in size, and data compression methods can facilitate efficient data storage, transfer, and utilization. Related workflows in geoscience modelling typically involve data generation (e.g. via weather forecasting), followed by compression, transfer, decompression, extraction of local subsets, and integrations into downstream tasks such as hydrological modelling. Recently, data-driven models have been introduced into geoscientific workflows with great success, allowing input data sources and types to be used in much more flexible ways than in classical numerical models, e.g. by directly using the compressed data. In this study, we therefore address the questions on 1) how various lossy compression algorithms compare in terms of general properties such as computational effort and their ability to extract spatial subsets directly from the compressed representation, and 2) how they compare in terms of compression efficiency and reproduction error. We focus on dimension reduction methods, as the resulting low-dimensional representations require fewer input channels in downstream data-driven models, thereby reducing computational costs and the risk of overfitting. We compare several dimension reduction methods in an application to five years of hourly air temperature and rainfall data, in the form of 2-d gridded spatial fields spanning a 100,000 km² domain in central Germany. The methods we compare are: Block Averaging, Principal Component Analysis, an Autoencoder, and the Ramer-Douglas-Peucker algorithm. We measure compression by the number of distinguishable objects in the compressed representation, rather than by file size, keeping in mind the potential use of the compressed data as input to data-driven models. This approach directly relates dimension reduction/compression to the number of input channels of a data-driven model, which is typically limited to avoid overfitting. Our results indicate that the methods differ substantially in terms of training and computational overhead, preservation of field-scale statistics, and the possibility to extract spatial subsets directly from the compressed representation. All methods demonstrated very good compression efficiency, with low reconstruction errors even for 99 % compression. For the spatially smooth temperature fields, Principal Component Analysis performed best, followed by the Autoencoder. For the spatially heterogeneous rainfall fields, the Autoencoder and the Ramer-Douglas-Peucker algorithm were most effective. The choice of the best method therefore depends on the specific application. This study provides a step towards a more efficient use of geoscientific data in downstream applications, either by directly using compressed data as input to data-driven models, or by efficiently extracting spatial subsets from compressed data, and using these subsets as input to local numerical models.
- 3 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-25
Open AccessComment on egusphere-2026-4514
Data sets in the geosciences continue to grow in size, and data compression methods can facilitate efficient data storage, transfer, and utilization. Related workflows in geoscience modelling typically involve data generation (e.g. via weather forecasting), followed by compression, transfer, decompression, extraction of local subsets, and integrations into downstream tasks such as hydrological modelling. Recently, data-driven models have been introduced into geoscientific workflows with great success, allowing input data sources and types to be used in much more flexible ways than in classical numerical models, e.g. by directly using the compressed data. In this study, we therefore address the questions on 1) how various lossy compression algorithms compare in terms of general properties such as computational effort and their ability to extract spatial subsets directly from the compressed representation, and 2) how they compare in terms of compression efficiency and reproduction error. We focus on dimension reduction methods, as the resulting low-dimensional representations require fewer input channels in downstream data-driven models, thereby reducing computational costs and the risk of overfitting. We compare several dimension reduction methods in an application to five years of hourly air temperature and rainfall data, in the form of 2-d gridded spatial fields spanning a 100,000 km² domain in central Germany. The methods we compare are: Block Averaging, Principal Component Analysis, an Autoencoder, and the Ramer-Douglas-Peucker algorithm. We measure compression by the number of distinguishable objects in the compressed representation, rather than by file size, keeping in mind the potential use of the compressed data as input to data-driven models. This approach directly relates dimension reduction/compression to the number of input channels of a data-driven model, which is typically limited to avoid overfitting. Our results indicate that the methods differ substantially in terms of training and computational overhead, preservation of field-scale statistics, and the possibility to extract spatial subsets directly from the compressed representation. All methods demonstrated very good compression efficiency, with low reconstruction errors even for 99 % compression. For the spatially smooth temperature fields, Principal Component Analysis performed best, followed by the Autoencoder. For the spatially heterogeneous rainfall fields, the Autoencoder and the Ramer-Douglas-Peucker algorithm were most effective. The choice of the best method therefore depends on the specific application. This study provides a step towards a more efficient use of geoscientific data in downstream applications, either by directly using compressed data as input to data-driven models, or by efficiently extracting spatial subsets from compressed data, and using these subsets as input to local numerical models.
- Open Access
- 16 October 2026
- Published by Copernicus GmbH in Hydrology and Earth System Sciences
- Vol. 30 (20) , 6339-6358
The need for uncertainty: why probabilistic LSTMs are key to improving flood predictions and enabling learned warning rules
Deterministic model predictions can struggle to adequately capture extreme events such as floods and droughts, which are of particular importance in hydrology. This limitation arises because deterministic models collapse the conditional discharge distribution to a single point estimate. Probabilistic modeling provides a way to address this issue by explicitly representing uncertainty and assigning non-zero probabilities to all possible outcomes, including extreme events, thereby capturing the entire conditional discharge distribution. Motivated by this perspective, we examine whether probabilistic Long Short-Term Memory (LSTM) models improve the representation of extreme events in rainfall–runoff simulations across Switzerland. The differing degree of flexibility in the models leads to different degrees and distinct patterns of miscalibration in the predictive distributions. For the extreme discharges (top 0.05 % of the marginal discharge distribution), the deterministic LSTM systematically underestimates 97 % of these events, whereas 83 % of these fall within the probabilistic model's 99th percentile prediction interval. Given the rarity of these events, representing them as outcomes near the upper tail of the predictive distribution is both expected and statistically consistent. Building on the additional information provided by probabilistic runoff predictions, we further show how they can be translated into actionable flood warnings using reinforcement learning. To this end, we introduce a Flood Risk Communication Agent (FRiCA) that operates on probabilistic runoff predictions and learns risk-aware decision rules for issuing warnings of varying intensity. Results indicate that the FRiCA outperforms simple fixed heuristics, such as issuing warnings based on the deterministic LSTM predictions or a fixed high quantile (e.g., the 99th percentile). The learnt behavior demonstrates the potential of reinforcement learning as a decision layer for improved flood risk communication, and motivates further exploration of better reward design and policy network optimization for context-dependent decision policies that adapt to varying hydrological and societal contexts.
- 16 October 2026
- Published by Copernicus GmbH in Hydrology and Earth System Sciences
- Vol. 30 (20) , 6339-6358
Open AccessThe need for uncertainty: why probabilistic LSTMs are key to improving flood predictions and enabling learned warning rules
Deterministic model predictions can struggle to adequately capture extreme events such as floods and droughts, which are of particular importance in hydrology. This limitation arises because deterministic models collapse the conditional discharge distribution to a single point estimate. Probabilistic modeling provides a way to address this issue by explicitly representing uncertainty and assigning non-zero probabilities to all possible outcomes, including extreme events, thereby capturing the entire conditional discharge distribution. Motivated by this perspective, we examine whether probabilistic Long Short-Term Memory (LSTM) models improve the representation of extreme events in rainfall–runoff simulations across Switzerland. The differing degree of flexibility in the models leads to different degrees and distinct patterns of miscalibration in the predictive distributions. For the extreme discharges (top 0.05 % of the marginal discharge distribution), the deterministic LSTM systematically underestimates 97 % of these events, whereas 83 % of these fall within the probabilistic model's 99th percentile prediction interval. Given the rarity of these events, representing them as outcomes near the upper tail of the predictive distribution is both expected and statistically consistent. Building on the additional information provided by probabilistic runoff predictions, we further show how they can be translated into actionable flood warnings using reinforcement learning. To this end, we introduce a Flood Risk Communication Agent (FRiCA) that operates on probabilistic runoff predictions and learns risk-aware decision rules for issuing warnings of varying intensity. Results indicate that the FRiCA outperforms simple fixed heuristics, such as issuing warnings based on the deterministic LSTM predictions or a fixed high quantile (e.g., the 99th percentile). The learnt behavior demonstrates the potential of reinforcement learning as a decision layer for improved flood risk communication, and motivates further exploration of better reward design and policy network optimization for context-dependent decision policies that adapt to varying hydrological and societal contexts.
- Open Access
- 8 October 2026
- Published by Copernicus GmbH in EGUsphere
- Vol. 2026, 1-25
Lossy compression by dimension reduction – Evaluation of methods applied to 2-d meteorological fields for data-based modeling
Data sets in the geosciences continue to grow in size, and data compression methods can facilitate efficient data storage, transfer, and utilization. Related workflows in geoscience modelling typically involve data generation (e.g. via weather forecasting), followed by compression, transfer, decompression, extraction of local subsets, and integrations into downstream tasks such as hydrological modelling. Recently, data-driven models have been introduced into geoscientific workflows with great success, allowing input data sources and types to be used in much more flexible ways than in classical numerical models, e.g. by directly using the compressed data. In this study, we therefore address the questions on 1) how various lossy compression algorithms compare in terms of general properties such as computational effort and their ability to extract spatial subsets directly from the compressed representation, and 2) how they compare in terms of compression efficiency and reproduction error. We focus on dimension reduction methods, as the resulting low-dimensional representations require fewer input channels in downstream data-driven models, thereby reducing computational costs and the risk of overfitting. We compare several dimension reduction methods in an application to five years of hourly air temperature and rainfall data, in the form of 2-d gridded spatial fields spanning a 100,000 km² domain in central Germany. The methods we compare are: Block Averaging, Principal Component Analysis, an Autoencoder, and the Ramer-Douglas-Peucker algorithm. We measure compression by the number of distinguishable objects in the compressed representation, rather than by file size, keeping in mind the potential use of the compressed data as input to data-driven models. This approach directly relates dimension reduction/compression to the number of input channels of a data-driven model, which is typically limited to avoid overfitting. Our results indicate that the methods differ substantially in terms of training and computational overhead, preservation of field-scale statistics, and the possibility to extract spatial subsets directly from the compressed representation. All methods demonstrated very good compression efficiency, with low reconstruction errors even for 99 % compression. For the spatially smooth temperature fields, Principal Component Analysis performed best, followed by the Autoencoder. For the spatially heterogeneous rainfall fields, the Autoencoder and the Ramer-Douglas-Peucker algorithm were most effective. The choice of the best method therefore depends on the specific application. This study provides a step towards a more efficient use of geoscientific data in downstream applications, either by directly using compressed data as input to data-driven models, or by efficiently extracting spatial subsets from compressed data, and using these subsets as input to local numerical models.
- 8 October 2026
- Published by Copernicus GmbH in EGUsphere
- Vol. 2026, 1-25
Open AccessLossy compression by dimension reduction – Evaluation of methods applied to 2-d meteorological fields for data-based modeling
Data sets in the geosciences continue to grow in size, and data compression methods can facilitate efficient data storage, transfer, and utilization. Related workflows in geoscience modelling typically involve data generation (e.g. via weather forecasting), followed by compression, transfer, decompression, extraction of local subsets, and integrations into downstream tasks such as hydrological modelling. Recently, data-driven models have been introduced into geoscientific workflows with great success, allowing input data sources and types to be used in much more flexible ways than in classical numerical models, e.g. by directly using the compressed data. In this study, we therefore address the questions on 1) how various lossy compression algorithms compare in terms of general properties such as computational effort and their ability to extract spatial subsets directly from the compressed representation, and 2) how they compare in terms of compression efficiency and reproduction error. We focus on dimension reduction methods, as the resulting low-dimensional representations require fewer input channels in downstream data-driven models, thereby reducing computational costs and the risk of overfitting. We compare several dimension reduction methods in an application to five years of hourly air temperature and rainfall data, in the form of 2-d gridded spatial fields spanning a 100,000 km² domain in central Germany. The methods we compare are: Block Averaging, Principal Component Analysis, an Autoencoder, and the Ramer-Douglas-Peucker algorithm. We measure compression by the number of distinguishable objects in the compressed representation, rather than by file size, keeping in mind the potential use of the compressed data as input to data-driven models. This approach directly relates dimension reduction/compression to the number of input channels of a data-driven model, which is typically limited to avoid overfitting. Our results indicate that the methods differ substantially in terms of training and computational overhead, preservation of field-scale statistics, and the possibility to extract spatial subsets directly from the compressed representation. All methods demonstrated very good compression efficiency, with low reconstruction errors even for 99 % compression. For the spatially smooth temperature fields, Principal Component Analysis performed best, followed by the Autoencoder. For the spatially heterogeneous rainfall fields, the Autoencoder and the Ramer-Douglas-Peucker algorithm were most effective. The choice of the best method therefore depends on the specific application. This study provides a step towards a more efficient use of geoscientific data in downstream applications, either by directly using compressed data as input to data-driven models, or by efficiently extracting spatial subsets from compressed data, and using these subsets as input to local numerical models.
- Open Access
- 22 June 2026
- Published by Copernicus GmbH
A Storyline-based Evaluation of Data-Driven Weather Models for Simulating the 2019 European Summer Heatwave under Future Warming Scenarios
Recently developed data-driven weather models demonstrate accuracy comparable to physical models while being computationally cheaper. Nonetheless, their efficacy in modeling extreme weather conditions in present day and their applicability in future warmer scenarios remains largely unknown. This study evaluates the performance of various global data-driven weather models in predicting the peak of the 2019 European summer heatwave, compared to the physics-based ICON model. The evaluation not only encompasses the present climate but also aims to assess the performance of these models in emulating this specific event in a +2, +3, and +4K warmer world relative to pre-industrial levels, following a storyline framework. Both data-driven and ICON models are initialized with global storyline simulations from the AWI-CM1 model for current and future warmer scenarios. To quantify the sensitivity to initial condition uncertainties, the random field perturbation technique is employed in the data-driven models. The findings indicate that the majority of data-driven models inadequately predict maximum temperatures in the heatwave core region under present climatic conditions as compared to ICON, exhibiting an underestimation of up to 2°C, even within an ensemble framework. These models are also assessed against ICON under future warmer climatic scenarios to determine their generalization capabilities outside the training distribution. The results indicate that these models can capture temperature amplification across climate scenarios to some extent, they underestimate the accelerated warming rate during the heat wave compared to ICON. This highlights both the potential and limitations of data-driven models for their applicability in future warmer scenarios and emphasizing the necessity to develop hybrid strategies combining physics with data-driven techniques to enhance predictive accuracy in a changing climate.
- 22 June 2026
- Published by Copernicus GmbH
Open AccessA Storyline-based Evaluation of Data-Driven Weather Models for Simulating the 2019 European Summer Heatwave under Future Warming Scenarios
Recently developed data-driven weather models demonstrate accuracy comparable to physical models while being computationally cheaper. Nonetheless, their efficacy in modeling extreme weather conditions in present day and their applicability in future warmer scenarios remains largely unknown. This study evaluates the performance of various global data-driven weather models in predicting the peak of the 2019 European summer heatwave, compared to the physics-based ICON model. The evaluation not only encompasses the present climate but also aims to assess the performance of these models in emulating this specific event in a +2, +3, and +4K warmer world relative to pre-industrial levels, following a storyline framework. Both data-driven and ICON models are initialized with global storyline simulations from the AWI-CM1 model for current and future warmer scenarios. To quantify the sensitivity to initial condition uncertainties, the random field perturbation technique is employed in the data-driven models. The findings indicate that the majority of data-driven models inadequately predict maximum temperatures in the heatwave core region under present climatic conditions as compared to ICON, exhibiting an underestimation of up to 2°C, even within an ensemble framework. These models are also assessed against ICON under future warmer climatic scenarios to determine their generalization capabilities outside the training distribution. The results indicate that these models can capture temperature amplification across climate scenarios to some extent, they underestimate the accelerated warming rate during the heat wave compared to ICON. This highlights both the potential and limitations of data-driven models for their applicability in future warmer scenarios and emphasizing the necessity to develop hybrid strategies combining physics with data-driven techniques to enhance predictive accuracy in a changing climate.
- Open Access
- 22 June 2026
- Published by Copernicus GmbH
Catchment-based post-processing of probabilistic weather forecasts for AI-supported flood prediction in Germany
Small and medium-size catchments respond rapidly to intense rainfall and are therefore especially challenging for flood forecasting, as short warning times coincide with large meteorological and hydrological uncertainties. Within the KI-HopE-De (AI-based flood prediction in small river basins in Germany) project, which aims to develop AI-based flood forecasting for small catchments, we investigate how probabilistic numerical weather prediction can be transformed into hydrologically more useful inputs for machine-learning-based runoff prediction. The project focuses on short-range forecasts up to 48 hours and on catchments smaller than 500 km².Our contribution addresses the meteorological side of this challenge through catchment-based post-processing of ensemble precipitation forecasts. We use hourly operational ICON-D2 ensemble precipitation forecasts with 3-hourly initializations and lead times up to 48 hours, aggregated from the native model grid to catchments. The resulting catchment-scale ensemble time series are post-processed using statistical and machine-learning-oriented methods designed to improve both calibration and hydrological relevance. In particular, we examine ensemble model output statistics (EMOS) and isotonic distributional regression (IDR) for marginal adjustment, and reconstruct temporal dependence using approaches such as ensemble copula coupling and the Schaake shuffle. This setup is designed to retain the benefits of probabilistic weather forecasts while generating physically and hydrologically more consistent forcing data for downstream flood-prediction models.The presentation will show how these post-processing strategies perform at two connected levels. First, we will compare post-processed precipitation forecasts against the raw ensemble using probabilistic verification metrics and event-based diagnostics, with particular emphasis on extremes. Second, we will assess whether improved meteorological inputs translate into improved flood forecasts in selected severe-flood cases. In this way, the study highlights not only whether post-processing improves rainfall forecasts, but also whether these improvements are relevant for downstream hydrological AI applications in small, fast-responding catchments.
- 22 June 2026
- Published by Copernicus GmbH
Open AccessCatchment-based post-processing of probabilistic weather forecasts for AI-supported flood prediction in Germany
Small and medium-size catchments respond rapidly to intense rainfall and are therefore especially challenging for flood forecasting, as short warning times coincide with large meteorological and hydrological uncertainties. Within the KI-HopE-De (AI-based flood prediction in small river basins in Germany) project, which aims to develop AI-based flood forecasting for small catchments, we investigate how probabilistic numerical weather prediction can be transformed into hydrologically more useful inputs for machine-learning-based runoff prediction. The project focuses on short-range forecasts up to 48 hours and on catchments smaller than 500 km².Our contribution addresses the meteorological side of this challenge through catchment-based post-processing of ensemble precipitation forecasts. We use hourly operational ICON-D2 ensemble precipitation forecasts with 3-hourly initializations and lead times up to 48 hours, aggregated from the native model grid to catchments. The resulting catchment-scale ensemble time series are post-processed using statistical and machine-learning-oriented methods designed to improve both calibration and hydrological relevance. In particular, we examine ensemble model output statistics (EMOS) and isotonic distributional regression (IDR) for marginal adjustment, and reconstruct temporal dependence using approaches such as ensemble copula coupling and the Schaake shuffle. This setup is designed to retain the benefits of probabilistic weather forecasts while generating physically and hydrologically more consistent forcing data for downstream flood-prediction models.The presentation will show how these post-processing strategies perform at two connected levels. First, we will compare post-processed precipitation forecasts against the raw ensemble using probabilistic verification metrics and event-based diagnostics, with particular emphasis on extremes. Second, we will assess whether improved meteorological inputs translate into improved flood forecasts in selected severe-flood cases. In this way, the study highlights not only whether post-processing improves rainfall forecasts, but also whether these improvements are relevant for downstream hydrological AI applications in small, fast-responding catchments.
- Open Access
- 14 March 2026
- Published by Copernicus GmbH
Probabilistic Benchmarks and Post-Processing for Data-Driven Weather Forecasting
In recent years, significant progress in machine learning technologies has enabled the development of various artificial intelligence weather prediction (AIWP) models, approaching, or even surpassing the skill of numerical weather prediction (NWP) models.However, despite these advancements, several important questions remain open. Most data-driven models primarily focus on deterministic point forecasts and lack the capability to generate probabilistic predictions, which, however, is crucial for optimal decision making and quantifying weather risk in applications. Further, while it has been widely demonstrated that physics-based NWP models substantially benefit from post-processing methods, which aim to correct systematic errors, the use of post-processing for data-driven weather models has not been explored in detail.Our overarching aim thus is to investigate the application of various post-processing techniques to potentially improve predictions, as well as to generate probabilistic forecasts from deterministic AIWP as well as NWP model outputs. We assess whether AI-based weather models benefit from post-processing to a similar extent as physics-based NWP, enabling a fair comparison between post-processed AIWP and NWP forecasts. The resulting post-processed AIWP forecasts also yield a relatively simple probabilistic benchmark for evaluating whether inherently probabilistic AIWP models deliver commensurate skill improvements given their increased computational cost.Experiments are based on the WeatherBench 2 framework, which provides a standardized archive of prominent AIWP as well as operational NWP model outputs. Specifically, we apply a suite of established statistical and machine learning post-processing methods to model outputs for the eight variables defined as headline scores (Z500, T850, Q700, WV850, T2M, WS10, MSLP, TP24hr) in the WeatherBench 2 framework, and systematically evaluate the effectiveness of these methods for improving deterministic and probabilistic forecasts.Results show that post-processed probabilistic forecasts can outperform the ensemble predictions from the European Centre for Medium-Range Weather Forecasts for shorter lead times of up to one week for selected variables, but the results vary across variables, lead times, post-processing methods and forecasting models.
- 14 March 2026
- Published by Copernicus GmbH
Open AccessProbabilistic Benchmarks and Post-Processing for Data-Driven Weather Forecasting
In recent years, significant progress in machine learning technologies has enabled the development of various artificial intelligence weather prediction (AIWP) models, approaching, or even surpassing the skill of numerical weather prediction (NWP) models.However, despite these advancements, several important questions remain open. Most data-driven models primarily focus on deterministic point forecasts and lack the capability to generate probabilistic predictions, which, however, is crucial for optimal decision making and quantifying weather risk in applications. Further, while it has been widely demonstrated that physics-based NWP models substantially benefit from post-processing methods, which aim to correct systematic errors, the use of post-processing for data-driven weather models has not been explored in detail.Our overarching aim thus is to investigate the application of various post-processing techniques to potentially improve predictions, as well as to generate probabilistic forecasts from deterministic AIWP as well as NWP model outputs. We assess whether AI-based weather models benefit from post-processing to a similar extent as physics-based NWP, enabling a fair comparison between post-processed AIWP and NWP forecasts. The resulting post-processed AIWP forecasts also yield a relatively simple probabilistic benchmark for evaluating whether inherently probabilistic AIWP models deliver commensurate skill improvements given their increased computational cost.Experiments are based on the WeatherBench 2 framework, which provides a standardized archive of prominent AIWP as well as operational NWP model outputs. Specifically, we apply a suite of established statistical and machine learning post-processing methods to model outputs for the eight variables defined as headline scores (Z500, T850, Q700, WV850, T2M, WS10, MSLP, TP24hr) in the WeatherBench 2 framework, and systematically evaluate the effectiveness of these methods for improving deterministic and probabilistic forecasts.Results show that post-processed probabilistic forecasts can outperform the ensemble predictions from the European Centre for Medium-Range Weather Forecasts for shorter lead times of up to one week for selected variables, but the results vary across variables, lead times, post-processing methods and forecasting models.
- Open Access
- 13 March 2026
- Published by Copernicus GmbH
Uncertainty quantification for data-driven weather prediction: From probabilistic forecasts to fair model comparisons
Artificial intelligence (AI)-based data-driven weather prediction (AIWP) models have experienced rapid progress over the last years. They achieve impressive results and demonstrate substantial improvements over state-of-the-art physics-based numerical weather prediction (NWP) models across a range of variables and evaluation metrics. However, most efforts in data-driven weather forecasting have been limited to deterministic, point-valued predictions, making it impossible to quantify forecast uncertainties, which is crucial in research and for optimal decision making in applications.I will present recent work on uncertainty quantification (UQ) methods in the context of data-driven weather prediction. The post-hoc use of UQ methods enables the generation of skillful probabilistic weather forecasts from a state-of-the-art deterministic AIWP model [1]. Further, by subjecting the deterministic backbone of physics-based and data-driven models post hoc to the same UQ technique, and computing the in-sample mean continuous ranked probability score of the resulting forecast, we propose a new measure that enables fair and meaningful comparisons of single-valued output from AIWP and NWP models, called potential continuous ranked probability score [2].References[1] Bülte, C., Horat, N., Quinting, J. and Lerch, S. (2025). Uncertainty quantification for data-driven weather models. Artificial Intelligence for the Earth System, in press. DOI:10.1175/AIES-D-24-0049.1[2] Gneiting, T., Biegert, T., Kraus, K., Walz, E.-M., Jordan, A. I., and Lerch, S. (2025). Probabilistic measures afford fair comparisons of AIWP and NWP model output. Preprint, arXiv:2506.03744. DOI:10.48550/arXiv.2506.03744
- 13 March 2026
- Published by Copernicus GmbH
Open AccessUncertainty quantification for data-driven weather prediction: From probabilistic forecasts to fair model comparisons
Artificial intelligence (AI)-based data-driven weather prediction (AIWP) models have experienced rapid progress over the last years. They achieve impressive results and demonstrate substantial improvements over state-of-the-art physics-based numerical weather prediction (NWP) models across a range of variables and evaluation metrics. However, most efforts in data-driven weather forecasting have been limited to deterministic, point-valued predictions, making it impossible to quantify forecast uncertainties, which is crucial in research and for optimal decision making in applications.I will present recent work on uncertainty quantification (UQ) methods in the context of data-driven weather prediction. The post-hoc use of UQ methods enables the generation of skillful probabilistic weather forecasts from a state-of-the-art deterministic AIWP model [1]. Further, by subjecting the deterministic backbone of physics-based and data-driven models post hoc to the same UQ technique, and computing the in-sample mean continuous ranked probability score of the resulting forecast, we propose a new measure that enables fair and meaningful comparisons of single-valued output from AIWP and NWP models, called potential continuous ranked probability score [2].References[1] Bülte, C., Horat, N., Quinting, J. and Lerch, S. (2025). Uncertainty quantification for data-driven weather models. Artificial Intelligence for the Earth System, in press. DOI:10.1175/AIES-D-24-0049.1[2] Gneiting, T., Biegert, T., Kraus, K., Walz, E.-M., Jordan, A. I., and Lerch, S. (2025). Probabilistic measures afford fair comparisons of AIWP and NWP model output. Preprint, arXiv:2506.03744. DOI:10.48550/arXiv.2506.03744
- Open Access
- 13 March 2026
- Published by Copernicus GmbH
Improving Flood Prediction and Warning through Probabilistic Deep Learning and Reinforcement Learning
Deterministic model predictions can struggle to adequately capture extreme events such as floods and droughts, which are of particular relevance in hydrology. This limitation arises because deterministic models collapse the conditional runoff distribution to a single point estimate. Probabilistic modeling provides a promising way to address this issue by explicitly representing uncertainty and assigning non-zero probabilities to a range of possible outcomes, including rare and extreme events, thereby capturing the full range of plausible hydrological responses. Motivated by this perspective, we investigate how long short-term memory (LSTM) based probabilistic models can be used for rainfall–runoff simulation across Switzerland. Overall, the probabilistic models show good calibration, although some miscalibration remains at the extremes. Differences between models mainly manifest in how uncertainty is distributed: some approaches produce narrower but lighter-tailed distributions, while others yield broader distributions with heavier tails. These trade-offs highlight that probabilistic models differ not only in sharpness but also in how they represent extreme outcomes. We also observe this trade-off in terms of the models’ single-point accuracy metrics. When evaluating the mean of the probabilistic predictions using the Nash–Sutcliffe efficiency (NSE), none of the probabilistic approaches outperform the deterministic LSTM in terms of average predictive accuracy. However, a clear advantage emerges when focusing on the tail of the discharge distribution. For the most extreme events (top 0.1% of the sorted discharge values), the deterministic LSTM underestimates more than 90% of observed values (since it provides estimates of an expectation), whereas probabilistic predictions can capture a substantially larger fraction of these extremes within their upper predictive bounds. Building on the additional information provided by probabilistic runoff predictions, we further show how such forecasts can be translated into discrete and actionable flood warnings using reinforcement learning. To this end, we introduce a Flood Risk Communication Agent (FRiCA) that operates on probabilistic runoff predictions and learns decision rules for issuing warnings of varying intensity. The FRiCA is implemented as an LSTM-based policy network and is trained by rewarding correct warning levels while penalizing the underestimation of flood severity. Results indicate that the FRiCA outperforms simple fixed heuristics, such as issuing warnings based on the predictive mean or a fixed high quantile (e.g., the 99th percentile). While this behavior already demonstrates the potential of reinforcement learning for improved flood risk communication, it also motivates future work toward more flexible and context-dependent decision strategies that adapt to varying hydrological and societal contexts.
- 13 March 2026
- Published by Copernicus GmbH
Open AccessImproving Flood Prediction and Warning through Probabilistic Deep Learning and Reinforcement Learning
Deterministic model predictions can struggle to adequately capture extreme events such as floods and droughts, which are of particular relevance in hydrology. This limitation arises because deterministic models collapse the conditional runoff distribution to a single point estimate. Probabilistic modeling provides a promising way to address this issue by explicitly representing uncertainty and assigning non-zero probabilities to a range of possible outcomes, including rare and extreme events, thereby capturing the full range of plausible hydrological responses. Motivated by this perspective, we investigate how long short-term memory (LSTM) based probabilistic models can be used for rainfall–runoff simulation across Switzerland. Overall, the probabilistic models show good calibration, although some miscalibration remains at the extremes. Differences between models mainly manifest in how uncertainty is distributed: some approaches produce narrower but lighter-tailed distributions, while others yield broader distributions with heavier tails. These trade-offs highlight that probabilistic models differ not only in sharpness but also in how they represent extreme outcomes. We also observe this trade-off in terms of the models’ single-point accuracy metrics. When evaluating the mean of the probabilistic predictions using the Nash–Sutcliffe efficiency (NSE), none of the probabilistic approaches outperform the deterministic LSTM in terms of average predictive accuracy. However, a clear advantage emerges when focusing on the tail of the discharge distribution. For the most extreme events (top 0.1% of the sorted discharge values), the deterministic LSTM underestimates more than 90% of observed values (since it provides estimates of an expectation), whereas probabilistic predictions can capture a substantially larger fraction of these extremes within their upper predictive bounds. Building on the additional information provided by probabilistic runoff predictions, we further show how such forecasts can be translated into discrete and actionable flood warnings using reinforcement learning. To this end, we introduce a Flood Risk Communication Agent (FRiCA) that operates on probabilistic runoff predictions and learns decision rules for issuing warnings of varying intensity. The FRiCA is implemented as an LSTM-based policy network and is trained by rewarding correct warning levels while penalizing the underestimation of flood severity. Results indicate that the FRiCA outperforms simple fixed heuristics, such as issuing warnings based on the predictive mean or a fixed high quantile (e.g., the 99th percentile). While this behavior already demonstrates the potential of reinforcement learning for improved flood risk communication, it also motivates future work toward more flexible and context-dependent decision strategies that adapt to varying hydrological and societal contexts.
- Open Access
- 28 February 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-33
Comment on egusphere-2026-469
Deterministic model predictions can struggle to adequately capture extreme events such as floods and droughts, which are of particular relevance in hydrology. This limitation arises because deterministic models collapse the conditional runoff distribution to a single point estimate. Probabilistic modeling provides a way to address this issue by explicitly representing uncertainty and assigning non-zero probabilities to a range of possible outcomes, including rare and extreme events, thereby capturing the full range of plausible hydrological responses. Motivated by this perspective, we examine whether probabilistic Long Short-Term Memory (LSTM) models improve the representation of extreme events in rainfall–-runoff simulations across Switzerland. Overall, the probabilistic models show good calibration, although some miscalibration remains for the extremes. Differences between models mainly manifest in how uncertainty is distributed: some approaches produce narrower and lighter-tailed distributions, while others yield broader distributions with heavier tails. These trade-offs highlight that probabilistic models differ not only in sharpness but also in how their calibration for rare events. We observe this tradeoff also in models' accuracy metrics. When evaluating the mean of the probabilistic predictions using the Nash–Sutcliffe efficiency (NSE), none of the probabilistic approaches outperform the deterministic LSTM in terms of average predictive accuracy. However, a clear advantage over the determinsitc models emerges when focusing on the tail of the discharge distribution. For the most extreme events (top 0.1 % of the discharge distribution), the deterministic LSTM underestimates more than 90 % of observed values (since it provides estimates of an expectation), whereas probabilistic predictions can capture a substantially larger fraction (67 %) of these extremes within their upper predictive bounds. Building on the additional information provided by probabilistic runoff predictions, we further show how they can be translated into actionable flood warnings using reinforcement learning. To this end, we introduce a Flood Risk Communication Agent (FRiCA) that operates on probabilistic runoff predictions and learns decision rules for issuing warnings of varying intensity. The FRiCA is implemented as an LSTM-based policy network and is trained by rewarding correct warning levels while penalizing the underestimation of flood severity. Results indicate that the FRiCA outperforms simple fixed heuristics, such as issuing warnings based on the predictive mean or a fixed high quantile (e.g., the 99th percentile). While this behavior already demonstrates the potential of reinforcement learning for improved flood risk communication, it also motivates further exploration of better reward design and policy network definition for context-dependent decision policies that adapt to varying hydrological and societal contexts.
- 28 February 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-33
Open AccessComment on egusphere-2026-469
Deterministic model predictions can struggle to adequately capture extreme events such as floods and droughts, which are of particular relevance in hydrology. This limitation arises because deterministic models collapse the conditional runoff distribution to a single point estimate. Probabilistic modeling provides a way to address this issue by explicitly representing uncertainty and assigning non-zero probabilities to a range of possible outcomes, including rare and extreme events, thereby capturing the full range of plausible hydrological responses. Motivated by this perspective, we examine whether probabilistic Long Short-Term Memory (LSTM) models improve the representation of extreme events in rainfall–-runoff simulations across Switzerland. Overall, the probabilistic models show good calibration, although some miscalibration remains for the extremes. Differences between models mainly manifest in how uncertainty is distributed: some approaches produce narrower and lighter-tailed distributions, while others yield broader distributions with heavier tails. These trade-offs highlight that probabilistic models differ not only in sharpness but also in how their calibration for rare events. We observe this tradeoff also in models' accuracy metrics. When evaluating the mean of the probabilistic predictions using the Nash–Sutcliffe efficiency (NSE), none of the probabilistic approaches outperform the deterministic LSTM in terms of average predictive accuracy. However, a clear advantage over the determinsitc models emerges when focusing on the tail of the discharge distribution. For the most extreme events (top 0.1 % of the discharge distribution), the deterministic LSTM underestimates more than 90 % of observed values (since it provides estimates of an expectation), whereas probabilistic predictions can capture a substantially larger fraction (67 %) of these extremes within their upper predictive bounds. Building on the additional information provided by probabilistic runoff predictions, we further show how they can be translated into actionable flood warnings using reinforcement learning. To this end, we introduce a Flood Risk Communication Agent (FRiCA) that operates on probabilistic runoff predictions and learns decision rules for issuing warnings of varying intensity. The FRiCA is implemented as an LSTM-based policy network and is trained by rewarding correct warning levels while penalizing the underestimation of flood severity. Results indicate that the FRiCA outperforms simple fixed heuristics, such as issuing warnings based on the predictive mean or a fixed high quantile (e.g., the 99th percentile). While this behavior already demonstrates the potential of reinforcement learning for improved flood risk communication, it also motivates further exploration of better reward design and policy network definition for context-dependent decision policies that adapt to varying hydrological and societal contexts.