OSA1.9

In weather applications, machine learning is emerging as an innovative technology with the potential to address many of the shortcomings of traditional modelling procedures. The trend is fostered by the growing availability of observational data, computational resources, and high-level software libraries. However, to ensure that machine learning can deliver on its promises, build trust, and eventually transition to become a reliable technology for production, it is also important to consider the technical and engineering challenges that arise when introducing machine learning in an operational environment.

MLOps is the set of practices that aims at deploying and maintaining machine-learning models in production continuously, reliably, and efficiently. At the core of MLOps is the idea that models can easily change, while the underlying workflows remain. In this sense, the emphasis is shifted from training a specific ML model to building an integrated ML system and to continuously operate it in production.

We will present our first experiences with MLOps for weather forecasting applications at MeteoSwiss. As an example, we will use a ML-based model for postprocessing NWP surface wind forecasts, as it covers the most common and relevant challenges, including the need for efficient data loading and manipulation, the monitoring and visualization of prediction quality, and the automation of model training and deployment pipelines. In this contribution, we aim at sharing our endeavor for MLOps best practices in the applied context of a national meteorological service with the hope to foster discussion and exchanges on the topic of machine learning operations for meteorological applications and their transition to the cloud.

How to cite: Nerini, D., Aznar, G., and Bhend, J.: Machine learning operations for weather applications, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-211, https://doi.org/10.5194/ems2022-211, 2022.

Ensemble weather predictions typically show systematic errors that need to be corrected via post-processing. While much research interest has been focused on univariate approaches, many practical applications such as energy forecasting, hydrological applications and air traffic management require accurate modeling of spatial, temporal, and inter-variable dependencies. Over the past years, a variety of two-step approaches where ensemble predictions are first post-processed separately in each margin and multivariate dependencies are restored via copula functions in a second step has been proposed to address this need [1]. However, these approaches share common limitations in that incorporating additional predictor variables beyond forecasts of the variable of interest is not possible in a straightforward manner in specifying the copula functions that govern the multivariate dependence structure, which makes it challenging to draw from substantial benefits that have recently been demonstrated in the context of univariate post-processing [2].

To address this challenge, we propose a novel data-driven one-step approach to multivariate ensemble post-processing based on conditional generative machine learning  which allows for obtaining multivariate probabilistic forecasts directly as output of a generative neural network while incorporating additional exogenous variables as predictors. In case studies on multivariate probabilistic forecasts of surface temperature and wind speed at observation stations in Germany, our conditional generative models show state-of-the-art forecast performance and advantages over benchmark approaches, for example by allowing for generating an arbitrary number of samples from the multivariate  forecast distributions. 

References
[1]  Lerch, S. et al. (2020) Simulation-based comparison of multivariate ensemble post-processing methods. Nonlinear Processes in Geophysics, 27: 349–371
[2] Rasp, S. and Lerch, S. (2018) Neural networks for post-processing ensemble weather forecasts. Monthly Weather Review, 146(11): 3885–3900

How to cite: Lerch, S. and Chen, J.: Generative machine learning methods for multivariate ensemble post-processing, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-195, https://doi.org/10.5194/ems2022-195, 2022.

Probabilistic predictions of precipitation call for rather sophisticated postprocessing approaches due to its low predictability, high spatio-temporal variability and highly positive skewness. Moreover, the large number of zeros makes the generation of physically realistic postprocessed forecast scenarios using standard approaches like ensemble copula coupling (ECC) rather difficult. In addition to classical statistical approaches, recently, machine learning based methods gained increasing popularity in the field of postprocessing of probabilistic weather forecasts.

In this study, we compare conditional generative adversarial network (cGAN) based postprocessing of daily precipitation with a quantile regression based approach. Recent publications on applying cGAN to precipitation forecasts have shown its potential to generate forecast scenarios that improve forecast skill and cannot be distinguished from observed data in terms of spatial structure (Harris et al., 2022; Price and Rasp, 2022). While we use ECC to generate physically realistic forecast scenarios from quantile regression, cGAN does not need any additional ECC steps. For training and verification, we use COSMO-E ensemble forecasts with a grid resolution of about 2 km over Switzerland and the corresponding CombiPrecip observations, which are a gridded blend of radar and gauge observations. Preliminary results obtained by using loss functions that are tailored to precipitation postprocessing as proposed by Harris et al. 2022 confirm the potential of cGAN for precipitation also for our study domain. Using cGAN, we aim to generate realistic looking forecast scenarios while also increasing forecast skill compared to COSMO-E. Furthermore, we provide a comparison of multivariate verification measures between COSMO-E, cGAN and quantile regression, which does increase forecast skill at the expense of relying on an additional ECC step to generate forecast scenarios. 

References: 

• Harris, L., McRae, A. T., Chantry, M., Dueben, P. D., & Palmer, T. N. (2022). A Generative Deep Learning Approach to Stochastic Downscaling of Precipitation Forecasts. arXiv preprint arXiv:2204.02028.

• Price, I., & Rasp, S. (2022). Increasing the accuracy and resolution of precipitation forecasts using deep generative models. arXiv preprint arXiv:2203.12297.

How to cite: Hemri, S., Bhend, J., Spirig, C., Nerini, D., Moret, L., Furrer, R., and Liniger, M. A.: Postprocessing of gridded precipitation forecasts using conditional generative adversarial networks and quantile regression, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-427, https://doi.org/10.5194/ems2022-427, 2022.

The problem of hydrologic modelling in large catchments has been addressed by conceptual physical-based models. Deep machine learning techniques such as Long Short-Term Memory (LSTM) networks have also proven to be effective for rainfall-runoff modelling. This is a promising approach to include a diversity of inputs to capture complex processes without handling the entire complexity, cost and difficulty that represent to include in conventional models. However, LSTM has not been extensively tested on large catchments and for seasonal forecasts in countries such as Germany. When a large catchment needs to be modelled, several questions arise, such as whether the performance of LSTM is suitable for rainfall-runoff without human intervention, to what extent we can rely on these results and whether the involved physical processes are appropriately represented also at higher spatial resolutions? The proposed study addresses these questions by developing a framework that employs daily meteorological observations and ancillary geographical information. First, by training a single LSTM to generate surface runoff in the study area. Secondly, to assess model performance, LISFLOOD is used as a benchmark spatially distributed semi-physical rainfall-runoff model. Finally, the performance of the model is evaluated against the Nash-Sutcliffe (NSE) and Kling-Gupta (KGE) efficiency criteria. The framework is illustrated using the Weser River Basin in Germany with a spatial resolution of 1 km and a daily time step. The proposed framework is expected to outperform calibrated physical-based models, to be suitable for seasonal forecasting, and to provide information for understanding the capabilities and limitations of the physics-based models and how the machine learning techniques take into account or are enforced to follow the physical laws.

How to cite: Espitia, E., Heidari, F., Lin, Q., Vischer, M., and Xoplaki, E.: Evaluating the performance of Long Short-Term Memory (LSTM) Networks for rainfall–runoff modelling in large catchments, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-541, https://doi.org/10.5194/ems2022-541, 2022.

During last year's summer storm season, we have introduced a precipitation nowcasting neural network MWNet and deployed it to operational use. The network tackles the nowcasting problem as a sequence to sequence prediction of radar echo, emphasizing high resolution and accuracy. We have conducted two quantitative experiments comparing MWNet 60 min forecasts to other available precipitation nowcasting models, using the metrics CSI and MSE. Both evaluations, over the domain of Denmark for years 2018 - 2020 and over the Czech Republic for the summer storm season of 2021, concluded in favor of our approach. However, we aim to improve MWNet capabilities further by focusing on severe weather nowcasting, the physical soundness of the predictions, and lead times longer than 60 min. Building on the advances in deep learning and its use in spatio-temporal forecasting, MWNet is based on the idea of disentangling physical dynamics from the residual factors. In this contribution, we consider improvements to the physical part of the network, its incorporation into the whole model, and the loss function used during training. Mainly, we are exploring the effect of implementing non-linear partial differential equations into the physical part, with various levels of hand-engineering equation terms. We analyze the impact on the dynamics learned by each part of the network and prediction quality for each setting. MWNet v1.2, based on the proposed architecture, will be operationally used and evaluated by meteorologists in Meteopress during the summer of 2022. This work aims to contribute to bridging the gap between machine learning and physical modeling in weather forecasting, alongside improving precipitation prediction.

How to cite: Choma, M., Bartel, J., and Šimánek, P.: Precipitation Nowcasting by Deep Physics-Constrained Neural Networks, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-391, https://doi.org/10.5194/ems2022-391, 2022.

Nowcasting severe weather is crucial not only to mitigate the effects of extreme weather events like storms and flash floods but also to support decision-makers on weather-dependent operations like aviation and outdoor events. Stunning pace of developments in Artificial Intelligence (AI) and increasing availability of high resolution data from different sensors motivate using AI in weather prediction, particularly in nowcasting due to its rapidly changing dynamics in short timescales.

Meteosat Third Generation (MTG) will greatly improve our capacity on nowcasting with its high resolution sensors on board. However, data-driven algorithms are needed to use the information stored within large volumes of MTG data for nowcasting in an efficient way. Therefore, we are developing AI based nowcasting algorithms fusing remote sensing data with ground based radar mosaics for nowcasting severe weather. 

As MTG data are not available yet (expected to be launched in late 2022), we use data from GOES satellites as the primary data source, which have comparable sensors with MTG. Specifically, we use the Storm Event ImageRy (SEVIR) dataset in the initial phase of the study which contains more than 10000 image sequences, 20 % of which contain storm events reported by NOAA. Each of these image sequences cover 384 x 384 km in space and 4-hour in time, containing three bands from the visible and infrared spectrums, and the lightning mapper data from GOES-16 with Vertically Integrated Liquid (VIL) mosaics derived from ground-based radar. We obtained promising results to reproduce spatial variations of VIL with Generative Adversarial Network (GAN) as a baseline. We are currently developing Recurrent Neural Network (RNN) based models to reproduce temporal variations of VIL to incorporate temporal information in GAN models. Furthermore, we are using European Weather Cloud, which not only provides a strong computation infrastructure but also fosters collaboration across projects. Development of efficient AI algorithms for nowcasting severe weather using GOES data will enable the opportunity to fully use MTG data on nowcasting severe weather.

How to cite: Küçük, Ç., Giannakos, A., Schneider, S., and Jann, A.: Towards a data-driven nowcasting of severe weather based on geostationary satellite data, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-343, https://doi.org/10.5194/ems2022-343, 2022.

Microwave radiometers (MWR) are moving into the focus of national and trans-national meteorological agencies which already operate or which intend to deploy MWR in network setups. The centralized processing of MWR data products within ACTRIS and the imminent integration of MWR into the EUMETNET E-Profile network are two prominent examples for this development. The developments within E-Profile correspond to efforts made by weather services towards directly assimilating MWR brightness temperature (TB) data.

At DWD we are evaluating data availability, quality, observation impact and operational sustainability of an MWR in a testbed setup, the so-called “Pilotstation”. In this framework first assimilation experiments of MWR TB at DWD were successful. The data assimilation (DA) requires a priori quality checks and an a priori detection of liquid water clouds. Currently the most frequent reason for rejecting data from DA is the suspected presence of clouds. Consequently, reliably identifying clouds without excessively rejecting clear-sky data is especially effective for increasing the availability of suitable data.

We will present a new approach for detecting the presence of liquid water clouds from the observed zenith TB. We employ a machine learning (ML) based algorithm which exploits the spectral signature and the variability of the observation. Using the CloudNet target classification as a reference we will demonstrate the overall performance. We will also highlight some issues faced and methods used during the development of this ML application.

TB observations at lower elevation angles increase the information content of the overall MWR observation. To ensure a proper interpretation of the TB the cloud detection scheme must be enhanced so that it also applies to off-zenith observations. We will present our approaches and first results at addressing this matter.

How to cite: Löffler, M., Knist, C., Görsdorf, U., Vural, J., and Löhnert, U.: New cloud detection method for a stand-alone ground based microwave radiometer, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-157, https://doi.org/10.5194/ems2022-157, 2022.

The transition to renewable energy sources such as solar energy has increased the interest in predicting cloud masks from remote sensing data. Even though deep learning methods have achieved great success on multiple meteorological tasks, only limited research has been conducted on nowcasting cloud masks based on high temporal and spatial resolution satellite data.

This study investigates forecasting cloud masks over Germany six frames into the future based on satellite images. We compare predictions between three deep learning architectures (ConvLSTM, U-Net, and MetNet) relative to two baseline models (optical flow and persistence). We train and evaluate our models using two years of the ICARE SAFNWC Cloud Mask dataset₁ , with a temporal resolution of 15 minutes per frame and a spatial resolution of 3×3 km per pixel. In our experiments we use a larger area of 256×256 pixels to predict the target area of size 128×128 pixels, leading to overall better performance compared to using an input size equal to the output size. Besides comparing different network architectures, we also investigate the effect of varying the temporal input size and output size for ConvLSTM. Finally, we examine the effect of adding more features (land/sea mask and elevation map) and changing the loss function.

In summary, we have performed a comprehensive study investigating cloud mask nowcasting using 70,000 spatially and temporally aligned data frames, examining three loss functions, six evaluation metrics, and three deep learning models.
During the presentation, we will highlight the main results from the study and present details of the model architectures, datasets, and how space and time affect the performance of the models.

1) Kniffka, Stengel, and Hollmann, “SEVIRI Cloud Mask Dataset - Edition 1.”

How to cite: Jungersen, M. E. M., Rasmussen, T. L., Nielsen, A. H., and Karstoft, H.: Cloud Mask Nowcasting over Germany Using Deep Learning, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-171, https://doi.org/10.5194/ems2022-171, 2022.

Atmospheric radiative transfer, which describes the evolution of radiation emitted by the Sun, the Earth's surface, clouds, and greenhouse gases, is an essential component of climate and weather modeling. In climate models, the transfer of radiation is approximated by parameterizations. Theoretically, however, with sufficient computing power, the electromagnetic radiation equations could be solved, but in practice this is out of reach.  The current operational radiative transfer solver in the Icosahedral Nonhydrostatic Weather and Climate Model (ICON) is ecRad, which, developed at ECMWF, is one of the most advanced available radiative transfer parameterizations.  It considers surface optics, gas optics, aerosol optics and cloud optics [1]. It is an accurate radiation parametrization but remains computationally expensive. Therefore, the radiation solver is usually not invoked at every time step and only runs on a reduced spatial grid, which can affect prediction accuracy, or only in a 1D setting without 3D transfer.

In this project, we are trying to develop an ecRad solver improved by machine learning to speed up the computation without loss of accuracy. Machine learning-based parametrizations would in general allow to fully replace existing sub-grid scale parameterizations, once trained from data. However, such parametrizations do not necessarily preserve essential physical quantities, which can lead to instabilities, model drifts or unphysical behavior as observed in [2] and [3].

We present here an emulation strategy, composed of three steps. First, we continue to call ecRad on a significantly coarser grid to predict the clear-sky radiation. We thereby use ecRad as a regularizer while reducing computation costs. Then, we interpolate the data on the full spatial grid using Gaussian processes. Finally, we predict the effect of the clouds on the radiation with random forests. The underlying idea is to avoid unphysical climate drifts and to support the generalization capabilities of the ML method.

Our first numerical experiments on an aqua planet simulation are promising. We hope to obtain a valuable outcome when considering more complex datasets with seasonality and realistic topography. Our final goal is to run a full ICON simulation with a machine learning enhanced ecRad parametrization, though the online performance remains open. There the clear-sky low resolution radiation field, computed with ecRad in the first part of our strategy, is expected to play a central role in model stability.

[1] Hogan, R. J., & Bozzo, A. (2018). A flexible and efficient radiation scheme for the ECMWF model. Journal of Advances in Modeling Earth Systems, 10, 1990-2008.

[2] Brenowitz, N. D., & Bretherton, C. S. (2018). Prognostic validation of a neural network unified physics parametrization. Geophysical Research Letters, 17, 6289–6298

[3] Brenowitz, N. D., and Bretherton, C. S. (2019). Spatially Extended Tests of a Neural Network Parametrization Trained by Coarse‐Graining, J. Adv. Model. Earth Syst., 11, 2728–2744

How to cite: Bertoli, G., Schemm, S., Ozdemir, F., Szekely, E., and Perez Cruz, F.: Building a physics-constrained, fast and stable machine learning-based radiation emulator, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-419, https://doi.org/10.5194/ems2022-419, 2022.

The seasonal climate forecasts at the Copernicus Climate Change Service (C3S) [1] provide longer-term predictions with several variables and forecast systems, serving as an outlook of weather statistics several weeks to months ahead. Hydrological models can assess natural hazards at regional scales when given high-resolution inputs of kilometer scales. Thus, to use the seasonal forecast ensembles for regional damage estimation, it is essential to increase the spatial resolution of these large sets of climate variables. Furthermore, the high-resolution data are adjusted to be physically consistent with the driving model in weather processes. Thanks to the latest advances in artificial intelligence, downscaling to high-resolution local data has become feasible. This study focuses on the comparison of three AI downscaling methods: multiple linear regression [2], artificial neural networks [3], and support vector machines [4]. In AI downscaling, the gridded data is used for model training and hyperparameter tuning. First, climate variables, including temperature, wind speed, precipitation, and solar radiation, are downscaled from 1° to 1 km. The downscaling results are then evaluated using statistical indicators compared with the historical daily station observations. Finally, the best-performing AI downscaling method is implemented to develop an early warning system to detect future climate extreme risks and their impacts on diverse economic activities in Germany.

[1] Seasonal forecast daily and subdaily data on single levels, Copernicus Climate Change Services, (2018). DOI: 10.24381/cds.181d637e.
[2] J. Bedia, J. Baño-Medina, M. N. Legasa, M. Iturbide, R. Manzanas, S. Herrera, A. Casanueva, D. San-Martín, A. S. Cofiño, J. M. Gutiérrez, Statistical downscaling with the downscaleR package (v3.1.0): Contribution to the VALUE intercomparison experiment, Geosci. Model Dev. 13 (3), 1711-1735 (2020). DOI: 10.5194/gmd-13-1711-2020.
[3] K. Ahmed, S. Shahid, S. B. Haroon, X. J. Wang, Multilayer perceptron neural network for downscaling rainfall in arid region: A case study of Baluchistan, Pakistan, J. Earth Syst. Sci. 124 (6), 1325-1341 (2015). DOI: 10.1007/s12040-015-0602-9.
[4] A. Goly, R. S. V. Teegavarapu, A. Mondal, Development and evaluation of statistical downscaling models for monthly precipitation 18 (18), 1-28 (2014). DOI: 10.1175/EI-D-14-0024.1.

How to cite: Lin, Q., Heidari, F., Espitia Sarmiento, E. F., Vischer, M., and Xoplaki, E.: Comparison of AI Downscaling Methods on C3S Seasonal Forecasts for Early Warning System Development, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-167, https://doi.org/10.5194/ems2022-167, 2022.

Heatwaves are extreme near-surface temperature events that can have substantial impacts on society and biodiversity. Moreover, the intensity, duration, and frequency of heatwaves are increasing at an accelerating rate as a consequence of climate change. Early Warning Systems can help to reduce the impact of heatwaves, as part of climate adaptation programs. However, state-of-the-art prediction systems can often not make accurate predictions of heatwaves more than two weeks in advance, which is required to take action and mitigate the impact of heatwaves. Here, we investigate central European forecasting of summer heatwaves on sub-seasonal timescales of several weeks using statistical and machine learning methods. As a first step, we select a set of atmospheric and surface predictors which are thought to have the largest impact on heatwave prediction based on previous studies and supported by a correlation analysis. Our findings show that at short lead times (1 week) near-surface temperature, 500hPa geopotential, precipitation, and surface soil moisture in central Europe are the most important predictors. At longer lead times (2—6 weeks), Mediterranean and North Atlantic sea surface temperatures, and the North Atlantic jet stream become the most relevant predictors. Secondly, we apply machine learning methods based on these predictors to forecast (1) summer temperature anomalies and (2) the probability of heat waves for 1—6 weeks lead time at weekly resolution. For each of these two types of forecasts (1) and (2) we use both a linear model and a Random Forest model. The performance of these models decays with lead time, as expected, but outperforms persistence and climatology at all lead times. Our machine learning models beat the European Centre for Medium-Range Weather Forecasts (ECMWF) model for lead times of 3 weeks and longer. We show that machine learning models can help extend the forecasting lead time of summer temperature anomalies and heat waves to sub-seasonal scales.

How to cite: Weirich Benet, E., Pyrina, M., Jiménez Esteve, B., Fraenkel, E., Cohen, J., and Domeisen, D.: Predicting Central European summer heatwaves with Machine Learning, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-233, https://doi.org/10.5194/ems2022-233, 2022.