Machine Learning and Computer Vision in Weather and Climate


Machine Learning and Computer Vision in Weather and Climate
Conveners: Jens Dittrich, Peter Düben, Richard Müller, Gordon Pipa, Bernhard Reichert, Dennis Schulze, Gert-Jan Steeneveld, Roope Tervo
Lightning talks
| Tue, 07 Sep, 14:00–15:30 (CEST)

Lightning talks: Tue, 7 Sep

Chairpersons: Bernhard Reichert, Roope Tervo
Machine Learning in NWP Postprocessing
Mark A. Liniger, Daniel Cattani, Benoit Crouzy, Daniele Nerini, Lionel Moret, and Christian Sigg

Machine Learning has a big potential for various tasks along the whole value chain of a national Met-Service. Indeed, many research groups, private and national weather services have started to explore the possibilities and first real-time operational implementations are in place already. However, the building up of the expertise is difficult, large amounts of data have to be made available in an efficient way and the necessary tools have special and demanding requirements concerning infrastructure and maintenance. Also, the transition from research results towards operational tools being operated in realtime is a particular challenge. Not least, trust from end-users must be built, while trying to avoid falling into the short-term hype trap.

In this presentation, we want to present some examples of machine-learning at MeteoSwiss that are in operational use or soon to be. This includes the use in a measurement system to identify pollen species, the quality control of meteorological observations, the postprocessing of numerical weather forecasts and the condensation of weather forecast information for the meteorologists. These examples have different characteristics and cover a wide range of applications, but also share some common properties. We want to juxtapose these properties with the incentives and conditions how machine learning methods are developed and employed in a more research oriented context like in academia. It turns out that an operational setup of machine learning has very different requirements than machine learning in a research context. The identification of these differences, but also the similarities, could help to understand the challenge of bringing research results into operation and how to alleviate this challenge in the future.

How to cite: Liniger, M. A., Cattani, D., Crouzy, B., Nerini, D., Moret, L., and Sigg, C.: Operational aspects of machine learning in a met-service, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-181, https://doi.org/10.5194/ems2021-181, 2021.

Daniele Nerini, Jonas Bhend, Christoph Spirig, Lionel Moret, and Mark Liniger

Hourly wind forecasts from numerical weather prediction models suffer from a range of systematic and random errors that are to a great extent related to limitations in the model grid resolution. To correct for such biases, statistical postprocessing and downscaling procedures are commonly applied so to leverage the information provided by automatic wind measurements at the surface. More recently, such techniques have been reformulated in a machine learning framework so to profit from the increased availability of data and computational resources. The results reported in the literature are promising and call for a serious evaluation of their potential for operational forecasting.

However, there remain several scientific and more applied challenges that need to be addressed before such methods can transition to real-world applications. One such challenge relates to the availability of multiple ensemble forecasts for the same point in time and space, which raises the question of how the information can be efficiently and optimally handled during postprocessing, so to provide added value to the end-user without adding technical debt to the operational system.

We propose an approach where a single deep learning model is trained to postprocess a combination of three ensemble forecasting systems, namely the high-resolution regional COSMO model with two configurations, and the ECMWF IFS ENS global ensemble forecasting system. We will show how the training is set up to provide a robust postprocessing model that can account for real time scenarios that include missing data and late model runs, while the quality of the forecasts remains comparable to a single-model approach. We found that the flexibility of the deep learning architecture translates into a robust automatic postprocessing solution that limits the maintenance burden and improves the system’s reliability.

How to cite: Nerini, D., Bhend, J., Spirig, C., Moret, L., and Liniger, M.: Seamless postprocessing of multi-model NWP surface wind forecasts with deep learning, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-277, https://doi.org/10.5194/ems2021-277, 2021.

Daan Scheepens, Katerina Hlavackova-Schindler, Claudia Plant, and Irene Schicker

The amount of wind farms and wind power production in Europe, on-shore and off-shore, increased rapidly in the past years. To ensure grid stability, omit fees in energy trading, and on-time (re)scheduling of maintenance tasks accurate predictions of wind speed and wind energy is needed. Especially for the prediction range of +48 hours up to 2 weeks ahead at least hourly predictions are envisioned by the users. However, these are either not covered by the high-resolution models or are on a spatial and temporal course scale. 

To address this as a first step we therefore propose a deep CNN based model for wind speed prediction  using the ECMWF ERA5 to train our model using at least seven wind-related temporal variables, i.e. divergence, geopotential, potential vorticity, temperature, relative vorticity, vertical wind velocity and horizontal wind velocity.

The input of the CNN is represented by  the 3-dim tensor (size of the 2-dim figures x time shots), one for each variable. The CNN  outputs the most probable of the six categories in which the wind speed will be during the following 96 hours, in 6h intervals. Different combinations of input data are investigated in terms of temporal input.

We analyse the influence of prediction range on the predicted category as well as the relevance of each of the wind-related variables in the prediction of this category.  The model will be tested and applied to the ECMWF IFS forecasts over Austria. The ensure a higher spatial and temporal resolution an additional step will be used for downscaling the CNN directly to a 1 km grid.

This work is performed as part of the MEDEA project, which is funded by the Austrian Climate Research Program.

How to cite: Scheepens, D., Hlavackova-Schindler, K., Plant, C., and Schicker, I.: A deep CNN model for medium-range spatio-temporal wind speed prediction for wind energy applications, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-238, https://doi.org/10.5194/ems2021-238, 2021.

Emy Alerskans, Joachim Nyborg, Morten Birk, and Eigil Kaas

It is a well-known fact that numerical weather prediction (NWP) models exhibit systematic errors, especially for near-surface variables. Reasons for this are, among other, the inability of these models to successfully handle sub-grid phenomena and shortcomings in the physical formulation of the model dynamics. Even though high-resolution regional NWP models usually have a spatial resolution of a few kilometers (or even finer) they generally exhibit local biases due to unresolved topography and obstacles. In order to obtain more local and site-specific forecasts post-processing methods can be used. Here, we have implemented a Transformer Neural Network model for post-processing 48-hour forecasts of 2 m temperature and relative humidity. The observational data used in this study consist of observations of 2 m air temperature and relative humidity from a network of private weather stations (PWSs). All in all, data from more than 1,000 locations are used. Forecast data from the Global Forecast System (GFS) model – such as temperature, relative humidity, wind speed and direction, radiation fluxes and upper level model fields – are also used as input to the model. The model is trained on 1.5 years of observational and forecast data and the performance is evaluated using an independent validation dataset of PWSs. We find that the Transformer post-processing model reduces the bias and standard deviation compared to the raw NWP forecast for a majority of stations. Furthermore, the model is validated on completely independent data from the Danish Meteorological Institute’s (DMI’s) observational network, where good results were obtained. Overall, the Transformer model produces forecasts that better match the locally observed weather.

How to cite: Alerskans, E., Nyborg, J., Birk, M., and Kaas, E.: A Transformer model for predicting near-surface temperature and relative humidity, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-198, https://doi.org/10.5194/ems2021-198, 2021.

Arnaud Mounier, Laure Raynaud, Lucie Rottner, and Matthieu Plu

The use of ensemble prediction systems (EPS) is challenging because of the huge information it provides. Forecasts from ensemble prediction systems (EPS) are often summarised by statistical quantities (ie quantiles maps). Although such mathematical representation is efficient for capturing the ensemble distribution, it lacks physical consistency, which raises issues for many applications of EPS in an operational context. In order to provide a physically-consistent synthesis of the French convection-permitting AROME-EPS forecasts, we propose to automatically draw a few scenarios that are representative of the different possible outcomes. Each scenario is a reduced set of EPS members.

To design a scenario synthesis, the procedure can be divided into two parts. A first step aims at extracting relevant features in each EPS member in order to reduce the problem dimensionality. Then, a clustering is done based on these features.

The originality of our work is to leverage the capacities of deep learning for the features extraction. For that purpose, we use a convolutional autoencodeur (CAE) to learn an optimal low-dimensional representation (also called latent space representation) of the input forecast field. In this work, the algorithm is developed to work on 1h-accumulated rainfall from AROME-EPS, with a focus on convective cases.

The CAE is trained on around 150 000 forecasts and its performance is evaluated based on the quality of the reconstructed input fields from the latent space. To examine the reconstruction quality, an object-oriented approach is used. CAE is also compared with the commonly-used principal component analysis (PCA). In a second part, different clustering methods (kmeans, HDBSCAN, …) are applied to EPS members in the latent space and evaluated using subjective and objective diagnostics.

How to cite: Mounier, A., Raynaud, L., Rottner, L., and Plu, M.: Weather scenarios from AROME-EPS forecasts using autoencoder and clustering methods, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-31, https://doi.org/10.5194/ems2021-31, 2021.

Machine Learning in Nowcasting
Richard Müller

Artificial Intelligence (AI) is nowadays a central key for the improvement of methods in many modern applications and research areas, e.g. autonomous driving, image / face recognition as well as system simulation and optimization. The AI success stories are increasing quite rapidly. However, AI is still largely underrepresented in weather services.

DWD was among the first European weather services who used modern AI methods within the scope of short term weather forecast. The respective research activities and applications are currently performed in close cooperation with the University of Saarbrücken and Mainz as well as the German Aerospace Center, whereby DWD acts mainly on implementation and steering of the research. Further, on an international level, DWD agreed recently on a closer cooperation with the new AI section of the South-Korean Meteorological Agency as well.

This presentation will provide an overview about the AI applications at DWD covering the following topics:

- The application of computer vision for the short term forecasting (nowcasting) of radiation, fronts, thunderstorms and precipitation based on satellite data.

- The development and application of machine learning for the early detection of thunderstorms as well as for the improvement of short term forecasts of mature Cbs.

- The operational usage of neuronal networks for the detection of volcanic ash (satellite data).

- AI concepts for a seamless integrated forecasting system.

However, AI, in particular machine learning, is also linked with handicaps. Thus the pros and cons of AI, will be discussed in comparison to classical methods. Within this scope, recent validation results will be considered as well.


How to cite: Müller, R.: AI for short term forecasts (nowcasting) at DWD, status and perspectives., EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-144, https://doi.org/10.5194/ems2021-144, 2021.

Francesco Zanetta and Daniele Nerini

Surface wind is an extremely difficult parameter to predict, particularly in the complex topography of the Alps. Due to several important processes happening at sub-kilometer scale, even high resolution Numerical Weather Prediction models such as COSMO-1 still present substantial biases. To address this, a wide range of statistical post-processing methods are used. Recently, methods based on Deep Learning have emerged as a new solution and are now actively developed at many weather services, including MeteoSwiss. At the same time, efforts are made to obtain accurate representations of surface wind speed up to a few hours ahead by integrating all available information in real-time, an approach known as nowcasting.

With the aim of seamlessly combining nowcasting and post-processing approaches for surface wind speed predictions, we developed a Deep Learning probabilistic post-processing model that is also able to integrate real time observations, and developed a new metric, the Similarity Index, for this purpose. The Similarity Index is a way to estimate the correlation of surface wind speed between two locations, based on their position and geomorphological setting, and can be used to choose the best available observation to be used at any point in space at any given time, and weigh that observation in a way that mimics geostatistical interpolation methods. The proposed methodology yields improved forecasts of wind speed where both systematic and random errors are reduced, thanks to the post-processing and nowcasting components respectively. In a second phase, we implemented a state- of-the-art explainability framework for machine learning, SHAP, and presented how it can be used to get insights into the model and build trust in the results.

How to cite: Zanetta, F. and Nerini, D.: Nowcasting of surface wind speed using probabilistic, explainable Deep Learning, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-331, https://doi.org/10.5194/ems2021-331, 2021.

Matej Choma, Jakub Bartel, Petr Šimánek, and Vojtěch Rybář

The standard for weather radar nowcasting in the Central Europe region is the COTREC extrapolation method. We propose a recurrent neural network based on the PredRNN architecture, which outperforms the COTREC 60 minutes predictions by a significant margin.

Nowcasting, as a complement to numerical weather predictions, is a well-known concept. However, the increasing speed of information flow in our society today creates an opportunity for its effective implementation. Methods currently used for these predictions are primarily based on the optical flow and are struggling in the prediction of the development of the echo shape and intensity.

In this work, we are benefiting from a data-driven approach and building on the advances in the capabilities of neural networks for computer vision. We define the prediction task as an extrapolation of sequences of the latest weather radar echo measurements. To capture the spatiotemporal behaviour of rainfall and storms correctly, we propose the use of a recurrent neural network using a combination of long short term memory (LSTM) techniques with convolutional neural networks (CNN). Our approach is applicable to any geographical area, radar network resolution and refresh rate.

We conducted the experiments comparing predictions for 10 to 60 minutes into the future with the Critical Success Index, which evaluates the spatial accuracy of the predicted echo and Mean Squared Error. Our neural network model has been trained with three years of rainfall data captured by weather radars over the Czech Republic. Results for our bordered testing domain show that our method achieves comparable or better scores than both COTREC and optical flow extrapolation methods available in the open-source pySTEPS and rainymotion libraries.

With our work, we aim to contribute to the nowcasting research in general and create another source of short-time predictions for both experts and the general public.

How to cite: Choma, M., Bartel, J., Šimánek, P., and Rybář, V.: High-Resolution Radar Echo Prediction with Machine Learning, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-159, https://doi.org/10.5194/ems2021-159, 2021.

Other Aspects of Machine Learning
Matti Kämäräinen, Kirsti Jylhä, Natalia Korhonen, and Otto Hyvärinen

Hot days, defined here as days exceeding the local 90th temperature percentile in summer months, pose an increasing threat to societies as summers warm along the climate. Therefore, an early warning of hot days and heat waves would be beneficial. To alleviate this need, we fit a convolutional neural network model to the global spatial distributions of the ERA5 reanalysis data to forecast the future number of hot days over the nearest 30-day period in Europe. 

A large set of potential input variable candidates were explored, including variables from the stratosphere and from the surface layers. Three-fold cross-validation was used to find the optimal subset to be used in forecasting. In addition to the input variables themselves, we use their temporal differences as predictors. Stepwise backward increasing of the amount of fitting data was applied to study the sensitivity of modelling to the number of fitting years. Finally, to emulate the real forecasting, time series hindcasting was applied by fitting a new model for each forecasted year, using only years prior to each year for fitting.

The target variable – the number of hot days during the nearest month – is extremely season-dependent. The non-linear forecasting model can take this into account, and both the grid cell based numbers of hot days and especially the mean numbers inside sub-regions show that the model is capable of reproducing the numbers. The skill, measured by the anomaly correlation coefficient, increases rapidly and constantly with an increasing number of fitting years. Interestingly, the skill curve does not level out, implying the model could still be enhanced by further increasing the fitting data.

How to cite: Kämäräinen, M., Jylhä, K., Korhonen, N., and Hyvärinen, O.: Forecasting monthly numbers of hot days in Europe with a convolutional neural network, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-90, https://doi.org/10.5194/ems2021-90, 2021.

Matthias Zech and Lueder von Bremen

Cloudiness is a difficult parameter to forecast and has improved relatively little over the last decade in numerical weather prediction models as the EMCWF IFS. However, surface downward solar radiation forecast (ssrd) errors are becoming more important with higher penetration of photovoltaics in Europe as forecasts errors induce power imbalances that might lead to high balancing costs. This study continues recent approaches to better understand clouds using satellite images with Deep Learning. Unlike other studies which focus on shallow trade wind cumulus clouds over the ocean, this study investigates the European land area. To better understand the clouds, we use the daily MODIS optical cloud thickness product which shows both water and ice phase of the cloud. This allows to consider both cloud structure and cloud formation during learning. It is also much easier to distinguish between snow and cloud in contrast to using visible bands. Methodologically, it uses the Unsupervised Learning approach tile2vec to derive a lower dimensional representation of the clouds. Three cloud regions with two similar neighboring tiles and one tile from a different time and location are sampled to learn lower-rank embeddings. In contrast to the initial tile2vec implementation, this study does not sample arbitrarily distant tiles but uses the fractal dimension of the clouds in a pseudo-random sampling fashion to improve model learning.

The usefulness of the cloud segments is shown by applying them in a case study to investigate statistical properties of ssrd forecast errors over Europe which are derived from hourly ECMWF IFS forecasts and ERA5 reanalysis data. This study shows how Unsupervised Learning has high potential despite its relatively low usage compared to Supervised Learning in academia. It further shows, how the generated land cloud product can be used to better characterize ssrd forecast errors over Europe.

How to cite: Zech, M. and von Bremen, L.: Understanding the relationship between clouds and surface downward radiation forecast errors with Unsupervised Deep Learning, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-471, https://doi.org/10.5194/ems2021-471, 2021.

Geoffrey Bessardon, Emily Gleeson, and Eoin Walsh

An accurate representation of surface processes is essential for weather forecasting as it is where most of the thermal, turbulent and humidity exchanges occur. The Numerical Weather Prediction (NWP) system, to represent these exchanges, requires a land-cover classification map to calculate the surface parameters used in the turbulent, radiative, heat, and moisture fluxes estimations.

The land-cover classification map used in the HARMONIE-AROME configuration of the shared ALADIN-HIRLAM NWP system for operational weather forecasting is ECOCLIMAP. ECOCLIMAP-SG (ECO-SG), the latest version of ECOCLIMAP, was evaluated over Ireland to prepare ECO-SG implementation in HARMONIE-AROME. This evaluation suggested that sparse urban areas are underestimated and instead appear as vegetation areas in ECO-SG [1], with an over-classification of grassland in place of sparse urban areas and other vegetation covers (Met Éireann internal communication). Some limitations in the performance of the current HARMONIE-AROME configuration attributed to surface processes and physiography issues are well-known [2]. This motivated work at Met Éireann to evaluate solutions to improve the land-cover map in HARMONIE-AROME.

In terms of accuracy, resolution, and the future production of time-varying land-cover map, the use of a convolutional neural network (CNN) to create a land-cover map using Sentinel-2 satellite imagery [3] over Estonia [4] presented better potential outcomes than the use of local datasets [5]. Consequently, this method was tested over Ireland and proven to be more accurate than ECO-SG for representing CORINE Primary and Secondary labels and at a higher resolution [5]. This work is a continuity of [5] focusing on 1. increasing the number of labels, 2. optimising the training procedure, 3. expanding the method for application to other HIRLAM countries and 4. implementation of the new land-cover map in HARMONIE-AROME.


[1] Bessardon, G., Gleeson, E., (2019) Using the best available physiography to improve weather forecasts for Ireland. In EMS Annual Meeting.Retrieved fromhttps://presentations.copernicus.org/EMS2019-702_presentation.pdf

[2] Bengtsson, L., Andrae, U., Aspelien, T., Batrak, Y., Calvo, J., de Rooy, W.,. . . Køltzow, M. Ø. (2017). The HARMONIE–AROME Model Configurationin the ALADIN–HIRLAM NWP System. Monthly Weather Review, 145(5),1919–1935.https://doi.org/10.1175/mwr-d-16-0417.1

[3] Bertini, F., Brand, O., Carlier, S., Del Bello, U., Drusch, M., Duca, R., Fernandez, V., Ferrario, C., Ferreira, M., Isola, C., Kirschner, V.,Laberinti, P., Lambert, M., Mandorlo, G., Marcos, P., Martimort, P., Moon, S., Oldeman,P., Palomba, M., and Pineiro, J.: Sentinel-2ESA’s Optical High-ResolutionMission for GMES Operational Services, ESA bulletin. Bulletin ASE. Euro-pean Space Agency, SP-1322,2012

[4] Ulmas, P. and Liiv, I. (2020). Segmentation of Satellite Imagery using U-Net Models for Land Cover Classification, pp. 1–11,http://arxiv.org/abs/2003.02899, 2020

[5] Walsh, E., Bessardon, G., Gleeson, E., and Ulmas, P. (2021). Using machine learning to produce a very high-resolution land-cover map for Ireland. Advances in Science and Research, (accepted for publication)

How to cite: Bessardon, G., Gleeson, E., and Walsh, E.: Using machine learning to produce a very high-resolution land-cover map for Ireland and beyond, EMS Annual Meeting 2021, online, 6–10 Sep 2021, EMS2021-53, https://doi.org/10.5194/ems2021-53, 2021.

Subsequent: Break out rooms to discuss each individual presentation

Supporters & sponsors