OSA1.9 | Machine Learning in Weather and Climate
Machine Learning in Weather and Climate
Conveners: Richard Müller, Gordon Pipa, Bernhard Reichert, Dennis Schulze, Gert-Jan Steeneveld, Roope Tervo
| Fri, 08 Sep, 09:00–13:00 (CEST)|Lecture room B1.05
| Attendance Thu, 07 Sep, 16:00–17:15 (CEST) | Display Wed, 06 Sep, 10:00–Fri, 08 Sep, 13:00|Poster area 'Day room'
Orals |
Fri, 09:00
Thu, 16:00
Artificial Intelligence (AI) is nowadays a central key for many modern applications and research areas, e.g. autonomous driving, image / face recognition as well as system simulation and optimization. Consequently, AI gains more and more importance also in weather and climate related sciences. This session focuses on machine learning techniques and aims at bringing together research with weather and climate related background with relevant contributions from computer sciences using these techniques.

Contributions from all kinds of machine learning studies in weather and climate on a wide range of time-scales are encouraged, including
• All kinds of postprocessing studies of Numerical Weather Prediction (NWP) forecasts
• Nowcasting studies, studies using satellite data, radar data, and observational weather data
• Seasonal forecast studies
• Climate related studies

These studies may e.g. deal with one or more of the fields
• Pre-processing of weather and climate data for machine learning purposes (e.g. forecasts from Numerical Weather Prediction (NWP) models, observational / satellite / radar data, etc.)
• Dimensionality reduction of weather and climate data, extraction of relevant features
• All kinds of supervised and unsupervised learning techniques
• Regression and classification tasks
• Artificial Neural Networks, Deep / Convolutional / Recurrent Neural Networks, LSTMs, Decision Trees, Support Vector Machines, Ensemble Learning and Random Forests, etc.
• Using cloud infrastructure to train and run AI-applications

Orals: Fri, 8 Sep | Lecture room B1.05

Chairpersons: Bernhard Reichert, Roope Tervo
Machine Learning in Weather Forecasting and NWP Postprocessing
Onsite presentation
Marvin Vincent Gabler, Jair Wuillaud, Hamid Taheri Shahraiyni, Daniela Neupert, Alexey Grigoryev, Rodrigo Almeida, Azamat Galimzhanov, Gabriel Martin Hernandez, Jordan Dane Daubinet, Nikoo Ekhtiari, Roan John Song, Peter Dudbridge, and Emrecan Tarakci

In this study, we present Jua’s Vilhelm, an innovative high-resolution AI-based global precipitation forecasting system. Vihelm exhibits significant improvements over existing state-of-the-art numerical models for prediction of binary precipitation (precipitation events) of up to 48 hours in advance, demonstrating superior performance when compared to the well-established Integrated Forecast System (IFS) model. Leveraging a combination of cutting-edge deep learning techniques and an in-depth understanding of physics, Vilhelm generates high-resolution hourly global predictions for surface parameters on a 1x1km grid. Beyond precipitation, the model is particularly adept at forecasts for a number of additional critical surface parameters including wind speed, direction and air temperature. This study focuses on its performance on precipitation prediction, notably the speed in which forecasts can be produced. The model runs in a matter of seconds, enabling the execution of hundreds of ensemble runs within a small number of minutes, an accomplishment previously unattainable. To assess the performance of Jua Vilhelm against Numerical Weather Prediction (NWP) models, we selected the IFS model of the European Centre for Medium-Range Weather Forecasts (ECMWF) as a suitable comparison. The ECMWF’s ERA5 reanalysis dataset was employed as a benchmark for evaluating the Vilhelm model on a global scale, using its full 0.25-degree resolution. We benchmarked both models at 6-hour intervals for multiple initial conditions for one year. Various metrics (Accuracy, Precision, Recall, Heidke Skill Score, F1 Score, True Negative Rate and False Alarm Rate) were used for comparative analysis between IFS precipitation forecasts and Vilhelm’s model. Vilhelm’s precipitation forecasts showed better performance than IFS over all metrics for prediction of precipitation events. Furthermore, precision, recall and P-R Curve of Vilhelm model for prediction of precipitation events were calculated using SYNOP (Surface Synoptic Observations) data as ground truth data. The results demonstrated the high performance of Vilhelm model. The results showed that development of Jua Vilhelm marks a significant advancement in the field of weather forecasting, offering unprecedented accuracy and speed.

How to cite: Gabler, M. V., Wuillaud, J., Taheri Shahraiyni, H., Neupert, D., Grigoryev, A., Almeida, R., Galimzhanov, A., Hernandez, G. M., Daubinet, J. D., Ekhtiari, N., Song, R. J., Dudbridge, P., and Tarakci, E.: A Novel High-Resolution AI-based Global Precipitation Forecasting System, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-355, https://doi.org/10.5194/ems2023-355, 2023.

Onsite presentation
Jonathan Demaeyer, Jonas Bhend, Sebastian Lerch, Cristina Primo, Bert Van Schaeybroeck, Aitor Atencia, Zied Ben Bouallègue, Jieyu Chen, Markus Dabernig, Gavin Evans, Jana Faganeli Pucer, Ben Hooper, Nina Horat, David Jobst, Janko Merše, Peter Mlakar, Annette Möller, Olivier Mestre, Maxime Taillardat, and Stéphane Vannitsem

Statistical postprocessing of forecasts produced by numerical weather prediction systems is an important part of modern weather forecasting systems.
Since the beginning of modern data science, numerous postprocessing methods have been proposed, and one of the questions that frequently arises is the relative performance of the methods for a given specific task.
However, a comprehensive, community-driven comparison of their relative performance is yet to be established. One of the main reasons for this lack is the difficulty of constructing a common comprehensive dataset that can be used to perform such intercomparison.
Here we introduce the first version of the EUPPBench, a dataset of time-aligned medium-range forecasts and observations over Central Europe, with the aim to facilitate and standardize the intercomparison of postprocessing methods. This dataset is publicly available [1], includes station and gridded data based on the ECMWF model forecasts [2]. The initial dataset is the basis of an ongoing project to establish a benchmarking platform for postprocessing of medium-range weather forecasts. We showcase a first benchmark of several methods for the adjustment of near-surface temperature forecasts and outline the future plans for the benchmark project activities.

[1] https://github.com/EUPP-benchmark/climetlab-eumetnet-postprocessing-benchmark

[2] Demaeyer, J., Bhend, J., Lerch, S., Primo, C., Van Schaeybroeck, B., Atencia, A., Ben Bouallègue, Z., Chen, J., Dabernig, M., Evans, G., Faganeli Pucer, J., Hooper, B., Horat, N., Jobst, D., Merše, J., Mlakar, P., Möller, A., Mestre, O., Taillardat, M., and Vannitsem, S.: The EUPPBench postprocessing benchmark dataset v1.0, Earth Syst. Sci. Data Discuss. [preprint], https://doi.org/10.5194/essd-2022-465, in review, 2023. 

How to cite: Demaeyer, J., Bhend, J., Lerch, S., Primo, C., Van Schaeybroeck, B., Atencia, A., Ben Bouallègue, Z., Chen, J., Dabernig, M., Evans, G., Faganeli Pucer, J., Hooper, B., Horat, N., Jobst, D., Merše, J., Mlakar, P., Möller, A., Mestre, O., Taillardat, M., and Vannitsem, S.: The EUPPBench postprocessing benchmark dataset, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-229, https://doi.org/10.5194/ems2023-229, 2023.

Onsite presentation
Kevin Höhlein and Tim Hewson

Application of a global downscaling method for 2m temperature is investigated, using forecasts of the ECMWF Integrated Forecast System. We work with a spatial resolution of 9km, which is destined to become the operational medium range ensemble resolution at ECMWF from June 2023. The downscaling technique is broadly based on Sheridan et al (Meteorological Applications, 2010), in which regression between model topographic elevation ("coarse grid") and model 2m temperature at an instant, in the vicinity of a gridbox in question, is the basis for lapse rate assignment for that gridbox. The rate so-derived is used to adjust temperatures onto a fine grid orography (e.g. 1km) for the said gridbox. Wherever, on the fine grid, elevation is out of the range of the coarse grid model topographic elevations in the vicinity then we use either extrapolation, for lower points, or model level data, for higher points. This overall approach should innately adapt to many meteorological scenarios - unstable, frontal, inversions, etc. - delivering much more weather-dependant lapse rate assignments for particular locations. The 2m temperatures so computed should verify much better than when using just a fixed rate assumption - e.g. -6.5K/km - as ECMWF does currently for it's very widely used meteogram products. The key attractions of this approach are: (i) the potential for large skill gains, (ii) that we exploit model "situational awareness" by indirectly incorporating the model's in-built physical processes, (iii) no training period or training data is needed, (iv) there is complete transparency for users because the lapse rate derivation is clearcut, and one can store lapse rate values on the grid in question.

In this presentation we will describe the approach and discuss the results of early investigations, using unstable and stable examples. The stronger and weaker aspects will be highlighted, discussing how to deal with coastal effects and addressing the all-important questions of search radius for defining 'the vicinity', and the regression format. One innate weaker point is the absence of any bias correction; a proposal for how this could also be built in in future, using weather-type dependant gridbox-scale bias correction from ECMWF's ecPoint post-processing method, for 2m temperature, will also be touched on.

How to cite: Höhlein, K. and Hewson, T.: 2m Temperature Downscaling with Dynamic Lapse Rates, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-301, https://doi.org/10.5194/ems2023-301, 2023.

Onsite presentation
Ivan Vujec and Iris Odak Plenković

NWP models have been a staple of modern weather forecasting for a long time. Although their skill is steadily improving, their limitations are still notable. An additional forecast improvement obtained by applying statistical post-processing for the locations where the measurements are available is thus a natural next step. Such an example of various NWPs improvements is achieved by the analog method.  For that reason, a potential post-processing contribution of the analog method is analyzed by verifying forecasts against the temperature measurements across the Republic of Croatia.


The raw NWP model used in this work is the ALADIN model with a 4 km horizontal resolution. The analyzed analog-based method includes weight optimization and the correction for more extreme values. Weight optimization aims to improve the process of similarity search by determining the ideal weight for each predictor variable, and prevents overfitting. The correction for more extreme values tends to improve forecasts on both ends of the temperature distribution by adding the correction factor if the starting NWP forecast exceeds the before-determined threshold.


The verification of the hourly temperature forecasts is performed using both the continuous and categorical approaches. In the categorical approach, the analysis is performed for both common and more extreme events. Considering the sensitivity of forecast accuracy on the terrain type, verification is also performed for different groups of stations. Results reveal the benefits of using the more advanced version of the analog method. Moreover, such an advanced analog method forecast outperforms raw NWP in most cases. When results are compared depending on the location, it is shown that the greater post-processing improvement can be seen in the continental than in the coastal area of Croatia.


Considering that the minimal and maximal daily values are a pragmatic way to provide forecasters with a quick and summarized overview, such forecasts are additionally examined. In addition to the statistical verification approach, a few interesting cases were also singled-out and analyzed more thoroughly. Finally, the amplitude of diurnal temperature variation seems to be more realistic for the analogs during heat or cold waves.

How to cite: Vujec, I. and Odak Plenković, I.: Analog-based post-processing of the operational temperature NWP in Croatia, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-441, https://doi.org/10.5194/ems2023-441, 2023.

Onsite presentation
Clément Brochet, Gabriel Moldovan, Laure Raynaud, and Matthieu Plu

Convection-permitting numerical weather prediction (NWP) models are a useful tool to forecast high-impact phenomena such as storms and heat waves. Ensemble prediction systems based on such models can then be used to quantify the associated forecast uncertainties. However, high-resolution ensemble forecasts come with a high computing cost; this limits the size of operational ensembles, and potentially reduces their relevance in critical situations.
In this work we propose a new and computationally efficient way to synthesize additional ensemble members of the kilometer-scale AROME model,based on deep generative models. For that purpose, state-of-the-art style-based generative adversarial networks (StyleGAN) and denoising diffusion probabilistic models (DDPM) are examined and compared on the joint generation of 10-meter wind and 2-meter temperature forecasts. We first show that these neural networks, once properly trained, create physically consistent, realistic, multivariate ensemble members. We then propose several methods to condition the generated members on the NWP model outputs 'of the day', in order to produce large hybrid physical/statistical ensembles at a small numerical cost. In particular, we show how to leverage the latent representations learnt by the networks to control the resulting ensemble statistics. A thorough evaluation of the resulting hybrid ensembles is proposed to assess their relevance and the new information they add with respect to small, pure NWP ensembles. Finally, we compare large, deep-learning-generated ensembles to a 875-member AROME forecast on a situation corresponding to a weather alert on the Mediterranean region, and evaluate the ability of generative approaches to provide an acceptable approximation of this large NWP ensemble.

How to cite: Brochet, C., Moldovan, G., Raynaud, L., and Plu, M.: Using state-of-the-art generative neural networks for high-resolution NWP ensemble emulation, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-402, https://doi.org/10.5194/ems2023-402, 2023.

Onsite presentation
Daniel Koser, Christoph Knigge, Björn-Rüdiger Beckmann, Dirk Zinkhan, Hermin Beumer-Aftahi, Benedikt Müller, Alexandra Melzer, Iris Breitruck, Stefan Seitz, and Helen Estrella

The Met4Airports project aims to apply methods of machine learning (ML) and artificial intelligence (AI) in order to provide predictions of the weather impact on planning and control parameters relevant for air traffic management (ATM), such as values of single and average flight delays as well as capacity values of runways and en-route airspace sectors.

For this purpose, air traffic data such as pre-scheduled flight lists and up-to-date time stamps from the A-CDM system (Airport Collaborative Decision Making) are combined with meteorological forecast data from nowcasting and numerical weather prediction models, with the ML models being trained on corresponding historical data sets.

While different approaches for modeling of the air traffic system regarding temporal discretization and sampling were investigated and extensive feature engineering and feature importance studies were conducted, a major challenge was the selection and optimization of appropriate ML model architectures to process the 2-dimensional data input from NowCastMix-Aviation (NCM-A) and the ICON-D2 model. Within this scope, it was found that comparatively simple multilayer perceptrons (MLPs), for whose data input the 2D NCM/ICON arrays are pooled in advance, show a better performance for all use cases than convolutional neural networks (CNNs), which are a staple of modern image processing.

Meteorological feature studies have been performed, aimed at determining the size of the relevant area around the airports, the required spatial resolution of the weather forecasting data and the relevant meteorological model prediction parameters. Additionally, also systematic hyperparameter studies were performed on all considered ML models.

Ultimately, the developed preliminary prototypes not only have shown to provide viable impact predictions on a set of thunderstorm days which were selected for historical process testing, but apparently also yield advantages over the already available information from the A-CDM system, which constitute a comparative baseline. A detailed insight on the validation and testing of the ML model predictions is given by Knigge et al. [1].  Furthermore, it was found that the temporal and spatial accuracy of the weather forecast provided by the ICON-D2 model at lead times of up to 24 hours is generally good enough for the impact prediction on ATM target parameters.

The project is funded by the German Federal Ministry for Digital and Transport (BMDV), coordinated by Deutscher Wetterdienst (DWD) and developed in close cooperation with the project partners ask – Innovative Visualisierungslösungen GmbH, Deutsche Flugsicherung (DFS), Fraport AG and Flughafen München GmbH (FMG). 

[1] Knigge et al. Testing and validation of forecasts for weather-induced operating restrictions in air traffic management based on Machine Learning models, OSA2.4 Reducing weather risks to transport: air, sea and land, submitted for EMS 2023.

How to cite: Koser, D., Knigge, C., Beckmann, B.-R., Zinkhan, D., Beumer-Aftahi, H., Müller, B., Melzer, A., Breitruck, I., Seitz, S., and Estrella, H.: Development and optimization of Machine Learning methods to predict weather-induced operating restrictions in air traffic management, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-274, https://doi.org/10.5194/ems2023-274, 2023.

Coffee break
Chairperson: Roope Tervo
Onsite presentation
Geoffrey Bessardon, Thomas Rieutord, and Emily Gleeson

Developing hectometric scale numerical weather prediction (NWP) models requires an accurate, high-resolution physiography dataset to represent surface heterogeneity in calculating surface-atmosphere exchanges. Land cover is one of the main components of physiography datasets. Unfortunately, the resolution of land cover datasets currently used in NWP is too coarse to fulfil the needs of hectometric NWP.

High-resolution remote sensing imagery and machine learning techniques have enabled the emergence of 10 m resolution global land cover maps such as  the Environmental Systems Research Institute 2020 (ESRI2020), European Space Agency (ESA) WorldCover, and the second generation Coordination of Information on the Environment land cover  (CLC+). These dataset labels are too generic to be directly implemented into NWP. Other high-resolution datasets can provide information about specific themes, such as the Copernicus dominant leaf type (DLT) 2018 and the Global map of Local Climate Zones, or national/regional information, such as the Swedish national land cover database 2018. While plenty of high-resolution physiography resources exist, none cover NWP needs.

This work aims at developing a high-resolution version of ECOCLIMAP-SG, the land cover map used operationally at Met Éireann, using supervised machine learning techniques. Supervised machine-learning techniques are widely used for land cover mapping but require a reference dataset. Here we propose to develop a reference dataset using multiple data sources and bagging multiple decision trees. We will complement this dataset with artificial (or synthetic) data using Geometric Synthetic Minority Oversampling Technique (G-SMOTE). Eventually, we will compare supervised machine-learning techniques such as convolutional neural networks or random forests.

How to cite: Bessardon, G., Rieutord, T., and Gleeson, E.: Developement of a land cover map for hectometric numerical weather prediction using machine learning, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-482, https://doi.org/10.5194/ems2023-482, 2023.

Machine Learning in Nowcasting and for Renewable Energy Applications
Onsite presentation
Kianusch Vahid Yousefnia, Tobias Bölle, Isabella Zöbisch, and Thomas Gerz

Thunderstorm forecasts with lead times of more than one hour usually rely on the post-processing of numerical weather prediction (NWP) data. Thanks to machine learning methods, this post-processing step has seen encouraging improvement in recent years. In this work, we introduce SALAMA, a feedforward neural network model for identifying thunderstorm occurrence in numerical weather prediction data. The model is trained on convection-resolving ensemble forecasts over Central Europe while lightning observations serve as ground truth. We believe that our work represents the first application of a neural network for thunderstorm forecasting to ensemble data with a fine resolution of only 2km. We solve a binary classification task: given only a set of pixel-wise input parameters that are extracted from NWP data and related to thunderstorm development, SALAMA infers the probability of thunderstorm occurrence. Particular emphasis is put on making the model reliable. We quantify classification skill through established scores from the meteorological and machine learning community and carefully estimate model uncertainty. For lead times up to eleven hours, we find a classification skill superior to classification based only on convective available potential energy. Varying the spatiotemporal criteria by which we associate lightning observations with NWP data, we estimate the advection speed of thunderstorms in the atmosphere and show that the time scale for skillful thunderstorm predictions increases linearly with the spatial scale of the forecast. All predictors entering our model are available in real time, which makes SALAMA readily available for operational use.

How to cite: Vahid Yousefnia, K., Bölle, T., Zöbisch, I., and Gerz, T.: Identification of thunderstorm occurrence in NWP forecasts using neural networks, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-10, https://doi.org/10.5194/ems2023-10, 2023.

Onsite presentation
Lukas Ivica, Juraj Bartok, Peter Sisan, Ivana Bartokova, Irina Malkin Ondik, and Ladislav Gaal

Fog is one of the severe meteorological phenomena, causing difficulties in transportation activities by air, road, or sea. Improving fog prediction methods is of great importance to society as a whole. Our study presents a fog forecast at the Poprad-Tatry Airport, Slovakia, where we used various machine learning algorithms (support vector machine, decision trees, k-nearest neighbors) to predict fog with visibility below 300 m for a lead time of 30 minutes. This research was carried out in the framework of the SESAR research project PJ.04-29.2. The novelty of the study is represented by the fact that beyond the standard meteorological variables as predictors, the forecast models also make use of information on visibility obtained through remote camera observations. The cameras help observe visibility using tens of landmarks at various distances and directions from the airport. The best-performing model achieved a score level of 0.89 (0.23) for the probability of detection (false alarm ratio). One of the most important findings of the study is that the predictor, defined as the minimum camera visibilities from eight cardinal directions, helps improve the performance of the constructed machine learning models in terms of an enhanced ability to forecast the initiation and dissipation of fog, i.e., the moments when a no-fog event turns into fog and vice versa. Camera-based observations help overcome the drawbacks of automated sensors (which predominantly measure points) and human observers (which have lower frequency observations but are more complex), and offer a viable solution for certain situations, such as the recent periods of the COVID-19 pandemic.

How to cite: Ivica, L., Bartok, J., Sisan, P., Bartokova, I., Malkin Ondik, I., and Gaal, L.: Machine Learning-Based Fog Nowcasting for Aviation with the Aid of Camera Observations, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-129, https://doi.org/10.5194/ems2023-129, 2023.

Onsite presentation
Petrina Papazek and Irene Schicker

Renewable Energy is gaining more importance in tackling globally growing energy demands. However, time-series obtained form production sites are often limited in time-horizon and also different in nature due to present technology in power plants. Due to their high resource demands high-resolution NWPs, such as AROME, can often not be fully stored long-term and change significantly in each update cycle, making them a challenging input for data hungry deep learning approaches. Still, in short- and medium range forecasts selected NWP parameters often improve forecast results of machine learning approaches.

In this study, we investigate how to setup a synthetically generated/extended training dataset based on diverse spatially and temporary heterogeneous inputs. For practical reasons, we concentrate on a solar power production case study. Data considered includes NWP models (e.g.: AROME, ECMWF), satellite data and products (e.g.:  CAMS), radiation time series from remote sensing, and observation time-series. In particular, we present a machine learning methodology on extending time-series from AROME in accordance with close observations.  Our synthetic data generator learns from long observation time series, and longer available but still coarse ECMWF. It yields both synthetic observations for missing data in the observation, synthetic production optimized for a relatively new established site, as well as a prolongation of numeric model data for studied renewable energy sites.

Results are evaluated by our deep learning framework (e.g.: LSTM) and cross-validation or achieved data of the data-set. The combination of real and synthetic data generally offers benefits in forecasting by more complex machine learning methods relying on sufficiently long training data-sets.

How to cite: Papazek, P. and Schicker, I.: Synthetic Data Generation for Deep Learning based Renewable Energy Forecasts, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-394, https://doi.org/10.5194/ems2023-394, 2023.

Onsite presentation
Nicolas Chea, Sylvain Cros, Sébastien Guillon, Jordi Badosa, Arttu Tuomiranta, and Martial Haeffelin

Forecasting solar energy is crucial for enabling its better integration into the energy mix. Short-term forecasts (up to 6h) have various applications, including grid integration and management, energy trading, energy storage management, and microgrid management. Photovoltaic (PV) production is highly variable and directly depends on solar irradiance reaching the Earth's surface. The stochastic component of solar irradiance variability is primarily related to its attenuation by the clouds.

Currently, cloud motion analysis methods applied to geostationary meteorological satellite images yield better results than Numerical Weather Prediction (NWP) models for short-term irradiance forecasting. Yet, their performance is limited in particular situations, such as convective or unstable cloud conditions. Deep Learning methods may overcome these limitations because of their ability to automatically extract complex spatiotemporal cloud patterns.

Prior to implementing a Deep Learning method, it is essential to analyse the data in order to understand its spatiotemporal characteristics. Clustering cloud patterns might allow us to better characterize cloud cover evolutions in distinct weather situations, thereby enabling the development of more accurate short-term solar energy forecasting models. Our research establishes a foundation for partition-based forecasts using multiple models, each model being adjusted to the distinct traits of each group.

In this work, we conduct a comparative analysis between traditional feature extraction techniques (e.g., statistical, textural, and temporal features) and Deep-Learning-based feature extraction (e.g., autoencoding) to evaluate their ability to discriminate specific cloud pattern evolutions (e.g., advections, convection, multi-layer, appearance/disappearance). Additionally, we assess and compare several clustering methods (e.g., K-Means and its variations, Hierarchical Clustering, Gaussian Mixture Model) using the extracted features as the input. Then, we evaluate the quality of the final partitioning using quantitative criteria and studying its physical consistency. Finally, we conduct an extensive study of the partitioning results, highlighting useful insights about the data such as the average duration of a given meteorological phenomenon. For this study, we use cloud albedo images derived from the HRV channel of the Meteosat Second Generation’s (MSG) SEVIRI instrument and focus on the Paris area.

The proposed feature extraction method enables us to perform clustering analysis that effectively distinguishes identifiable meteorological situations and cloud pattern evolutions. This is preliminary work for the development of a partition-based Deep Learning model for short-term solar energy forecasting, with the goal of addressing complex and diverse forecast scenarios using multiple models.

How to cite: Chea, N., Cros, S., Guillon, S., Badosa, J., Tuomiranta, A., and Haeffelin, M.: Cloud patterns clustering in geostationary satellite imagery: Investigating traditional and Deep-Learning-based feature extraction techniques for solar energy forecast, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-434, https://doi.org/10.5194/ems2023-434, 2023.

Onsite presentation
Hadrien Verbois, Vadim Becquet, Yves-Marie Saint-Drenan, Benoit Gschwind, and Philippe Blanc

Accurate estimations of Surface Solar Irradiance (SSI) are of high interest in domains as varied as climatology, solar energy, architecture, and agriculture. SSI estimations derived from meteorological satellites enable continuous spatial and temporal coverage and have thus become an important source of information standardly used for the planning, operation, and forecast of the production of PV power systems. To infer SSI from satellite images is, however, not straightforward; since the eighties, multiple satellite-based retrieval approaches have been proposed, from the earlier cloud index methods to physically based ones. Recent approaches are emerging based on machine learning (ML), inferring a direct data-driven model between images acquired from satellite and SSI ground measurements. Although only a few such works have been published, their practical efficiency has already been questioned. The objective of this paper is not to propose a new ML-based method but to better understand the promises and limitations of this new coming family of methods.

To do so, we implement simple multi-layer-perceptron models with different training datasets of satellite-based radiance measurements from Meteosat Second Generation (MSG) with collocated SSI ground measurements. To test the model's ability to generalize in time and space, we use different locations and time periods for training and testing. How we allocate measurement stations to each group is also a crucial factor. To understand our model behavior, we study two setups. In the first setup, stations are randomly assigned to each group, resulting in distinct but spatially interlaced training and test stations. In the second setup, we enforced strict and large geographical separation, allowing us to evaluate the model's performance in locations outside its training area. In both cases, the performance of the ML-based retrieval model is compared to that of the operational CAMS radiation service (CRS), which is based on Heliosat-4, a state-of-the-art physical retrieval model.

Our results show that the data-driven model’s performance can be much better than CRS but is very dependent on the training set, raising problems of generalization. Indeed, in the first setup, the ML model has a Root Mean Square Error (RMSE) almost 20% lower than CRS, but in the second training setup – when training and test stations are geographically separated, CRS RMSE is only 4% higher. Perhaps more critically, in the first setup, the ML model RMSE is lower or comparable to that of CAMS for all test stations but in the setup enforcing geographical separation, the ML model underperforms dramatically for several test locations.

ML models have great potential for satellite retrieval, but their inability to generalize in certain configurations could be critical and hinder their deployment to regions with sparse measurement networks. A hybrid approach combining data-driven and physical models seems to be of interest for further research activities.

How to cite: Verbois, H., Becquet, V., Saint-Drenan, Y.-M., Gschwind, B., and Blanc, P.: Promises and limitations of machine-learning-based methods for satellite retrieval of solar surface irradiance, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-460, https://doi.org/10.5194/ems2023-460, 2023.

Machine Learning for Seasonal to Sub-Seasonal Scales
Onsite presentation
Nina Horat and Sebastian Lerch

Even though sub-seasonal weather forecasts are crucial for the decision making in many sectors including agriculture, public health and renewable energy production, sub-seasonal predictions are hardly used due to their low skill. We propose several post-processing approaches based on convolutional neural networks (CNNs) to improve and calibrate sub-seasonal forecasts from numerical weather prediction (NWP) models. All proposed methods work directly with the spatial NWP forecasts and are therefore able to retain spatial dependencies in the forecasts. Moreover, these methods have the potential to exploit the predictive information in the spatial structure of the NWP forecasts.

The proposed post-processing models use forecast fields of multiple meteorological variables as input, and produce global probabilistic tercile forecasts for biweekly aggregates of temperature and precipitation for weeks 3-4 and 5-6, as commonly done in sub-seasonal to seasonal prediction. The model architectures and the training strategy are optimized to deal with the low signal-to-noise ratio in sub-seasonal forecasts and the limited amount of training data. Half of the tested architectures use the well-known UNet architecture specifically designed for image segmentation. The remaining architectures are based on a standard CNN as typically used for image classification, with the difference that it estimates coefficient values for a set of basis functions to provide spatial predictions.

All post-processing methods improve precipitation and temperature forecasts for both lead times. Improvements increase with lead time and are larger for precipitation, for which the NWP forecasts show no skill. Our best performing model is based on the UNet architecture and trained directly on the global NWP forecasts. The post-processed forecasts are substantially less sharp than the respective probabilistic forecasts from ECMWF for all tested methods. We demonstrate that the post-processed forecasts are well calibrated in contrast to the ECMWF benchmark using a calibration simplex, a reliability diagram for probabilistic tercile forecasts. Thus, our proposed post-processing methods are able to derive a reliable uncertainty estimate based only on ensemble mean NWP forecasts.

How to cite: Horat, N. and Lerch, S.: Deep learning for post-processing global probabilistic forecasts on sub-seasonal time-scales, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-211, https://doi.org/10.5194/ems2023-211, 2023.

Onsite presentation
Yang Liu, Bart Schilperoort, Peter Kalverla, Sem Vijverberg, Jannes van Ingen, Sarah Alidoost, Stefen Verhoeven, and Dim Coumou

Reliable S2S forecasts remain a huge scientific challenge. Only for specific ‘windows of predictability’, skillful forecasts are possible, in an otherwise largely unpredictable future. Due to a number of successes in S2S forecasting, the interest in machine learning (ML) is growing fast. However, we argue there is a need for more standardization, consensus on best practices, higher efficiency, and higher reproducibility. Typical S2S ML use-cases, such as (1) pure statistical forecasting based on observations, (2) transfer learning, and (3) post-processing of dynamical model ensembles, require a large coding and preprocessing effort. Such experiments are not trivial to set up, and without sufficient experience and expertise there is a large risk of improper cross-validation and/or improper and non-standard verification. 

Driven by the need for a reliable tool to integrate expert knowledge and artificial intelligence, we are developing two python packages to tackle the scientific challenge of (sub) seasonal (S2S) forecasting. s2spy is a high level python package designed to handle and optimize the entire data-driven forecasting workflow. Lilio is an advanced calendar system for resampling timeseries into training and target data for machine learning. Our aim is to make ML workflows more transparent and easier to build, and to facilitate standardization and collaboration across the S2S community. This also contributes to higher reproducibility and works towards a wider acceptance of standards and best practices. We will present our vision and the capabilities of our package, show-casing that we can build a model from raw climate data up to verification in only a few lines of code. 

How to cite: Liu, Y., Schilperoort, B., Kalverla, P., Vijverberg, S., van Ingen, J., Alidoost, S., Verhoeven, S., and Coumou, D.: Data-driven s2s forecasting with s2spy & lilio, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-501, https://doi.org/10.5194/ems2023-501, 2023.

Posters: Thu, 7 Sep, 16:00–17:15 | Poster area 'Day room'

Display time: Wed, 6 Sep 10:00–Fri, 8 Sep 13:00
Chairperson: Roope Tervo
Pedro Jeuris, Giuliano Andrea Pagani, and Mirela Popa

The phenomenon of fog poses a considerable risk to transportation by land, air and sea. The resulting lower visibility conditions on land can cause road accidents which in turn can result in economical damages, material damages and in the worst cases human casualties too. This raises the need of a wide-spread fog monitoring and warning system to prevent these sort of dreadful events. Specialized equipment to detect fog exists, but is expensive, which makes a large-scale use infeasible. Instead we propose the use of an infrastructure already available, such as traffic monitoring cameras, in combination with Deep Learning Computer Vision techniques. This approach has several advantages such as 1) no specialized or additional equipment is necessary, 2) predictions can be visually verified by an operator before taking a decision to raise a fog related warning and 3) cheap scalability in terms of adding cameras where and when needed. In this work, two well established computer vision models (i.e., ResNet and ViT) are tested on this task and explainable AI techniques are applied to better understand which parts of the input are important for the classification task. Furthermore, we extend these computer vision models to a multimodal setting by adding meteorological variables to increase the classification performance. To our knowledge this is the first work combining visual inputs and meteorological data for the task of fog detection. Additionally, given the limited amount of annotated images and the high effort required to label them, we also explored the application of self-supervised methods to increase the performance. We obtained promising results, even though we did not manage to surpass the accuracy of the supervised approach. The best setting for training achieves an accuracy of 91% and an F1 score of 85% on the test set.

How to cite: Jeuris, P., Pagani, G. A., and Popa, M.: WTF: Where There's Fog, Detecting Low Visibility Conditions on Highways Using Multimodal Computer Vision, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-9, https://doi.org/10.5194/ems2023-9, 2023.

Markus Rosenberger, Manfred Dorninger, and Martin Weißmann

Clouds play an important role in atmospheric processes like radiation and moisture transport and knowing the currently occurring cloud type can give the observer insights into the dynamics going on in the atmosphere. However, the number of human cloud observers is rather decreasing than increasing and thus efficient methods for automated cloud classification are sought as a way to continuously gather such information. Moreover, such a method is also less subjective than any human observer. Machine Learning methods and especially Convolutional Neural Networks (CNNs) have shown exceptionally good results in a broad range of image classification tasks. Despite the fact that machine learning has already been used to classify clouds from satellite data during the last few decades, the field of cloud classification from conventional RGB pictures taken at the Earth's surface is rather untouched. Although the WMO's classification scheme is defined based on visual properties of clouds only visible from below their undersurface. Hence, in this work CNNs are trained to discriminate between up to 30 cloud types from conventional RGB pictures. The used data set consists of all sky panorama images taken by the cloud observation system of the Department of Meteorology and Geophysics at the University of Vienna during the period 2016-2019 as well as from 2022 onwards. Ground truth labels are taken from hourly operational observations at the station Vienna Hohe Warte. The fact that more than one cloud type can occur at once makes this task a multi class multi label classification problem and even harder to solve. However, once trained properly the resulting algorithm can be used together with an ordinary camera to classify clouds with a high temporal resolution. Possible applications may be, e.g. model verification or to efficiently monitor the current status of the weather as well as its short-time evolution. First results will be shown.

How to cite: Rosenberger, M., Dorninger, M., and Weißmann, M.: Training CNNs for cloud classification from RGB pictures, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-137, https://doi.org/10.5194/ems2023-137, 2023.

Haolin Liu, Jimmy C.H. Fung, and Alexis K.H. Lau

Precise quantification of precipitation is crucial for effective planning and minimizing property damage or loss of human life caused by extreme weather events, especially under the rapidly changing climate. Currently, quantitative precipitation forecasting (QPF) in numerical weather prediction (NWP) models rely heavily on parameterization schemes for microphysics, boundary layers, cumulus, etc., rather than directly solving physical-based governing equation sets to predict fundamental variables such as temperature, wind speed, and humidity. These parameterization schemes introduce significant uncertainties in precipitation forecasting due to the limited knowledge of precipitation processes, which bottlenecks the performance of precipitation forecasting in NWP models.

To overcome this challenge, we propose a deep learning model based on Vision-Transformer that directly ingests fundamental meteorological variables solved by NWP models as predictors and maps them quantitatively to the precipitation map from a satellite-merged precipitation product. In this study, we conducted Weather Research and Forecasting (WRF) model simulations at 27km grid resolution for five years from 2017 to 2021 over China and the southeast region of Asia, and we used simulation results for the wettest season from June to September in 2017-2019 as training data, while validating and testing the model performance on data from 2020 and 2021. The deep learning model aims to circumvent uncertainties in physical parameterization schemes, which are due to the incomplete understanding of physical processes, and directly reproduce the high-resolution satellite rainfall observation product, the Climate Prediction Center morphing method (CMORPH) data.

Our evaluation results on the test dataset show that the deep learning model effectively extracts features from meteorological variables, leading to improved precipitation skill scores of 21.7%, 60.5%, and 45.5% for light rain, moderate rain, and heavy rain, respectively, on an hourly basis. We also evaluate two case studies under different synoptic conditions and show promising results in estimating heavy precipitation during strong convective precipitation events. Overall, the proposed deep learning model can provide vital insights for capturing precipitation-triggering mechanisms and enhancing precipitation forecasting skills. Additionally, we discuss the sensitivities of the fundamental meteorological variables used in this study, training strategies, and performance limitations.

How to cite: Liu, H., Fung, J. C. H., and Lau, A. K. H.: Enhancing quantitative precipitation estimation in the NWP using a deep learning model, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-27, https://doi.org/10.5194/ems2023-27, 2023.

Inchae Na, Sojung An, Taejin Oh, Jiyeon Jang, and Wooyeon Park

This study utilized a convolutional neural network (CNN) architecture based on Convlstm and Trajgru models with CBAM method to predict rainfall. The radar and satellite data collected and preprocessed during summer seasons from 2019 to 2022 were used along with additional auxiliary data including longitude, latitude, and terrain information to perform deep learning-based rainfall prediction.
CBAM method is used to focus attention on specific spatial and channel relationships. Convlstm and Trajgru are neural network architectures designed to process spatiotemporal data.
Assuming that atmospheric flows move at a maximum speed of 72 km/hour, to predict based on information six hours ahead of time, a spatial context of 432 km in all directions, which includes a 512 km X 512 km domain, was used for training.
In order to train and validate the deep neural network, the produced radar data from the Korean Meteorological Administration was compared and evaluated. 
The model was trained to predict rainfall for the next 0-360 minutes using a combination of continuous GK2A satellite data and radar images captured every 30 minutes starting from 2 hours before the current time.
The predictive models trained with different loss functions, including Huber loss, Cosh loss, and a combination of MSE and MAE with adjustable ratios, were compared based on their performance in predicting the thresholded rainfall intensities of 0.1, 1, 4, and 9 mm/hr, as evaluated by the critical success index (CSI). 
In addition, the rainfall intensities were classified into eight categories using a segmentation technique, and the performance of the models was analyzed accordingly.

How to cite: Na, I., An, S., Oh, T., Jang, J., and Park, W.: Deep learning model for six hour rainfall prediction, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-181, https://doi.org/10.5194/ems2023-181, 2023.

Christian Berndt and Markus Schultze

Object-based cell detection and tracking algorithms provide a useful tool for analyzing current and past storm properties and movement. KONRAD3D, an algorithm which recently became operational at DWD, uses radar scans of different elevation angles to derive 3-dimensional cell objects with specific properties.  Nowcasting of storm position by displacing it using the current cell movement seems to be straightforward, however, predicting the life cycle or specific storm properties such as lifetime, maximum future severity, and hail occurrence and size is difficult. We aim at analyzing the potential of machine learning techniques (ML), in particular random forest and gradient boosting, to provide these predictions using KONRAD3D cell properties in combination with NWP and lightning activity data.

We perform a recalculation of KONRAD3D for a 6-year time period, where we considered several other data sources to compute specific storm environment and attributes. For instance, NWP data from the ICON-EU model is used to characterize the convective environment, while lightning data is used quantify electrical activity. Next, we filter resulting storm detections regarding our three prediction tasks using ML:

(1) Maximum expected future cell severity: all storms are considered.

(2) Longevity of quasi-stationary cells: only storms with a minimum speed below a threshold are considered.

(3) Hail size: all storms are considered.

For tasks 1 and 2, we analyze the prediction performance for different storm ages, i.e. cell attributes at initial detection, after 15 min, and 30 min during the life cycle are used as predictors for ML. Maximum hail size (task 3) is estimated using individual cell detections. In order to achieve a final comparison of methods, we perform a 10-fold cross validation

Longevity of quasi-stationary storm events is difficult to predict since there many events with a short duration and only very few with a long duration. The distribution of maximum expected severity is less skew and therefore prediction scores are better. Random forest and gradient boosting are preferred over artificial neural networks, since decision tree methods are easier and faster to train and also provide better cross validation scores. In case of maximum future severity, a substantial improvement compared to a simple persistence forecast is achieved. ML-based prediction of maximum hail size delivers reasonable results and is able to outperform classical methods, such as MESH. However, prediction uncertainty remains high in most cases and needs to be quantified in order to generate meaningful predictions.

How to cite: Berndt, C. and Schultze, M.: ML-based prediction of storm properties using KONRAD3D cell objects, lightning data, and environment conditions from NWP, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-451, https://doi.org/10.5194/ems2023-451, 2023.

Marco Di Giacomo, Lorenzo Giuliano Papale, Francesca Marcucci, Mario Papa, Raffaele Golino, and Fabio Del Frate

In the present study, the research activities aiming at investigating the use of modern Machine Learning algorithms for the early detection of convective systems are shown.
The study conceived and carried out at the Italian Air-Force Meteorological Centre, co-funded in the framework of the EUMETNET-SRNWP-EPS Project, focuses on the post-processing of NWP model output to provide to the forecasters an improved Decision Support System specifically designed for aviation hazards and severe weather phenomena, such as thunderstorms. In detail, the NWP fields generated from operational Limited Area Models (COSMO-IT, 2.2 km horizontal resolution) running at CNMCA (Centro Nazionale Meteorologia e Climatologia Aerospaziale, Italian AirForce Met Service) and ECMWF, were processed and fed, as input features, to a properly designed decision tree. Different targets were defined to train the ML model, starting from independent observation datasets, which include radar, lightning and satellite-derived observations. The products mentioned above and the input features were co-located in space and time to assess the correlation between the signal (the information content of NWP products) and the weather phenomenon (e.g., convective instability). 
The ML tool operates in near real-time mode for the prediction/test phase and offline mode for the training/validation phase. It runs on the European Weather Cloud (EWC), the cloud-based collaboration platform for European meteorological application development. In this context, the customizable computational capabilities of the EWC allowed us to manage the extensive dataset and reduce the model training time. Then the trained models were validated by considering external observations (i.e., the METAR dataset, which provides airports' messages concerning the main meteorological phenomena and parameters observed by weather stations, such as thunderstorms. Moreover, some case studies will be shown and, as a benchmark of the potential improvement, a comparison with classical (multi-variate, static-threshold decision-tree) post-processing applications, routinely used by the Meteorological Watch Office of the Centre, is included in the evaluation study.

How to cite: Di Giacomo, M., Giuliano Papale, L., Marcucci, F., Papa, M., Golino, R., and Del Frate, F.: A ML approach for Post-processing of NWP model on the European Weather Cloud, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-489, https://doi.org/10.5194/ems2023-489, 2023.

Bu-Yo Kim, Yun-Kyu Lim, and Joo Wan Cha

Particulate matter (PM) increases traffic accidents and mortality rates in the short term, and poses a long-term threat to public health, such as cardiovascular and respiratory diseases, deteriorating human health. Therefore, real-time monitoring and prediction are crucial for improving air quality, and efforts should be made to regulate and reduce the emissions of air pollutants at the national or regional level. In this study, tree-based machine learning algorithms (RF, XGB, LGB) were used with local data assimilation and prediction system (LDAPS) meteorological forecast data (four times a day, hourly forecast for 36 hours) to predict PM10 and PM2.5 in Seoul, South Korea. Seoul, being a densely populated megacity, experiences high local emissions of air pollutants and significant influx of high-concentration pollutants from neighboring countries and desert areas. Therefore, monitoring and prediction of PM concentrations are of utmost importance in Seoul. The prediction results were compared with the observed data of PM10 and PM2.5 from air-quality measurement station in Seoul. The results showed that LGB method among the tree-based ML algorithms was the most suitable for PM prediction, with hourly R2=0.83~0.86 and daily R2=0.996. These results showed that the prediction performance was significantly higher than the chemical transport model prediction results, with R2 being more than 0.2 higher, and the prediction performance was superior especially in cases of high-concentration PM. Therefore, the high-accuracy PM prediction based on machine learning presented in this study can be useful for air quality monitoring and improvement.


Acknowledgments: This work was funded by the Korea Meteorological Administration Research and Development Program “Research on Weather Modification and Cloud Physics” under Grant (KMA2018-00224).

How to cite: Kim, B.-Y., Lim, Y.-K., and Cha, J. W.: PM10 and PM2.5 prediction in Seoul, South Korea using LDAPS meteorological data and tree-based machine learning, EMS Annual Meeting 2023, Bratislava, Slovakia, 4–8 Sep 2023, EMS2023-175, https://doi.org/10.5194/ems2023-175, 2023.