Forecasting the weather

Forecasting the weather, in particular severe and extreme weather has always been the most important subject in meteorology. This session will focus on recent research and developments on forecasting techniques, in particular those designed for operations and impact oriented. Contributions related to nowcasting, meso-scale and convection permitting modelling, ensemble prediction techniques, and statistical post-processing are very welcome.
Topics may include:
 Nowcasting methods and systems, use of observations and weather analysis
 Mesoscale and convection permitting modelling
 Ensemble prediction techniques
 Ensemble-based products for severe/extreme weather forecasting
 Seamless deterministic and probabilistic forecast prediction
 Post-processing techniques, statistical methods in prediction
 Use of machine learning, data mining and other advanced analytical techniques
 Impact oriented weather forecasting
 Presentation of results from relevant international research projects of EU, WMO, and EUMETNET etc.

Co-organized by NH1
Convener: Yong Wang | Co-conveners: Jing Chen, Ken Mylne, David Richardson, Guido Schröder
vPICO presentations
| Mon, 26 Apr, 15:30–17:00 (CEST)

vPICO presentations: Mon, 26 Apr

Chairperson: Yong Wang
Gabriele Messori, Stephen Jewson, and Sebastian Scher
Users of meteorological forecasts are often faced with the question of whether to make a decision now based on the current forecast or whether to wait for a later and hopefully more accurate forecast before making the decision. Imagine that you are the organiser of an event planned for Saturday. If the weather conditions at the start of the event are unsuitable then the event will have to be cancelled, leading to various expenses. Daily weather forecasts are available in the run-up to the event and you need to use them to decide whether to cancel in advance or not. Cancelling early could lead to only small cancellation charges, while cancelling shortly before leads to larger charges. Both sets of cancellation charges are lower than the potential loss due to last-minute cancellation on Saturday, and this leads to a nuanced set of decisions around when and whether to cancel. The general mathematical framework for understanding decisions of this type has been studied extensively, both in meteorology and in other fields such as economics. In order to understand our problem of whether to decide now or wait for the next forecast, we consider a special case of this general framework, that is also an extension of the well-known cost-loss model. We find that within this extended cost-loss model, the question of whether to decide now or wait depends on probabilities of probabilities. We develop a decision algorithm which we apply to real forecasts of temperature, and find that the algorithm leads to better decisions in most settings relative to three simpler alternative decision-making schemes. Our results have implications for the additional kinds of information that weather and climate forecasters could produce to facilitate good decision making based on their forecasts.

How to cite: Messori, G., Jewson, S., and Scher, S.: Decide now or wait for the next forecast? Testing a decision framework using real forecasts and observations, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-75,, 2020.

Anastase Charantonis, Vincent Bouget, Dominique Béréziat, Julien Brajard, and Arthur Filoche

Short or mid-term rainfall forecasting is a major task with several environmental applications such as agricultural management or flood risks monitoring. Existing data-driven approaches, especially deep learning models, have shown significant skill at this task, using only rainfall radar images as inputs. In order to determine whether using other meteorological parameters such as wind would improve forecasts, we trained a deep learning model on a fusion of rainfall radar images and wind velocity produced by a weather forecast model. The network was compared to a similar architecture trained only on radar data, to a basic persistence model and to an approach based on optical flow. Our network outperforms by 8% the F1-score calculated for the optical flow on moderate and higher rain events for forecasts at a horizon time of 30 minutes. Furthermore, it outperforms by 7% the same architecture trained using only rainfall radar images. Merging rain and wind data has also proven to stabilize the training process and enabled significant improvement especially on the difficult-to-predict high precipitation rainfalls. These results can also be found in Bouget, V., Béréziat, D., Brajard, J., Charantonis, A., & Filoche, A. (2020). Fusion of rain radar images and wind forecasts in a deep learning model applied to rain nowcasting. arXiv preprint arXiv:2012.05015

How to cite: Charantonis, A., Bouget, V., Béréziat, D., Brajard, J., and Filoche, A.: Fusion of rain radar images and wind forecasts in adeep learning model applied to rain nowcasting, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-11990,, 2021.

Seongsim Yoon and Hongjoon Shin

It is important to utilize various hydrological and weather information and accurate real-time forecasts to understand the hydrological conditions of the dam in order to make decisions of dam operation. In particular, due to rainfall concentrated in a short period of time during the flood season, it is necessary to plan the exact amount of dam discharge using real-time rainfall forecasting information. Compared to the ground rain gauge network, the radar has a high resolution of time and space, which enables the continuous expression of rainfall, which is very advantageous for very short-term prediction. Especially, In particular, the radar is capable of three-dimensional observation of the atmosphere, which has an advantage in understanding the vertical development and structure of clouds and rainfall, which can be used to observe torrential rain in the dam basin and to anticipate future rainfall intensity changes, rainfall movement and duration time. This study aims to develop a suitable radar-based very short-term rainfall prediction technique and to produce rainfall prediction information of the dam basin for stable dam operation and water disaster prevention. The radar-based rainfall prediction in this study is to be performed using a convolutional deep neural network with the 8 years weather radar data of the Korea Meteorological Administration. And, we select rainfall cases with high rainfall intensity and train the deep neural network to ensure the accuracy of flood season rainfall prediction. In addition, we intend to perform the accuracy evaluation with extrapolation-based rainfall prediction results for the dam basin.


This work was supported by KOREA HYDRO & NUCLEAR POWER CO., LTD (No. 2018-Tech-20)

How to cite: Yoon, S. and Shin, H.: Very short-term radar rainfall prediction using deep neural network for hydropower dam operation, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-1054,, 2021.

Stephen Moseley

Knowledge of the expected precipitation phase is crucial for mitigating the impacts snow and ice on national infrastructure. This is sensitive to the altitude of the modelled forecast grid point which varies between models.

The IMPROVER project aims to blend probabilistic model variables from different models. This presentation describes the approach used to standardise the phase change levels of falling precipitation from the Met Office UK and Global models over the high-resolution UK domain.

The method uses wet-bulb temperature profiles to identify the surface where snow changes to sleet and sleet changes to rain, interpolates these surfaces through model orography and below sea level, then extracts the predicted phase at the altitude of the standard high-resolution UK orography. This is performed for each model realization to maintain the multivariate connection between precipitation and precipitation phase.

The precipitation phase discriminators are used to categorise precipitation rate and accumulation probability data into rain, sleet and snow phases which in turn inform a categorical most-likely weather code.

We present results from a one-month trial using data from February 2020 comparing the weather code forecasts with site observations across the UK.

How to cite: Moseley, S.: A precipitation phase discriminator for IMPROVER, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-15184,, 2021.

Dominique Brunet and John Rafael Ranieses Quinto

The phase of falling precipitation can have a large societal impact for both hydrology (snow storage, rain-on-snow events), meteorology (snowstorms, freezing rain) and climate (snow albedo feedback). In Canada, many surface weather stations report precipitation information in the form of total precipitation (liquid-equivalent), but very few weather stations directly report snow. Thus, precipitation phase must be inferred from ancillary data such as temperature and moisture. Each scientific community has developed its own tool for the determination phase in the absence of direct observations: from simple rules based on air temperature, dew point temperature or wet bulb temperature to sophisticated microphysics schemes passing by methods based on the discrimination of features extracted from vertical temperature profiles. With the recent advances of machine learning, there is an opportunity to investigate another set of methods based on deep neural networks.

Using ERA5 and ERA5-Land model re-analyses as the reference, we trained several recurrent neural networks (RNN) on vertical profiles of temperature and moisture to infer the snow fraction – the ratio of solid precipitation to total precipitation. Since precipitation phase (solid, liquid or mixed) was not directly available in the model re-analysis, we defined it using two thresholds: snow fraction of less than 5% for liquid, snow fraction of more than 95% for solid phase, and mixed phase for everything in between. The best performing neural network for regressing snow fraction is found to be a Gated Recurrent Unit (GRU) RNN using profiles up to 500 hPa above the surface of both temperature and relative humidity. A slight decrease in performance is observed if profiles up to 700 hPa are used instead. A feature experiment also reveals that the performance is significantly better when using both temperature and moisture profiles, but it does not really matter what type of moisture observations are used (either dew point spread, wet bulb temperature or relative humidity). For classifying precipitation phase, the balanced accuracy is over 90%, clearly outperforming the implementation of Bourgouin’s method used operationally in part of Canada. Compared with the K-Nearest Neighborhood (KNN) method trained on surface observations only, it is seen that the greatest gain in performance for GRU-RNN is when the surface temperature is close to zero degrees Celsius.

These preliminary results indicate the great potential of the proposed algorithm for determining snow fraction and precipitation phase in the absence of direct observations. The proposed algorithm could potentially be used for inferring snow fraction and precipitation phase in several applications such as (1) precipitation analysis for forcing hydrological models, (2) weather nowcasting, (3) weather forecast post-processing and (4) climate change impact studies.


How to cite: Brunet, D. and Quinto, J. R. R.: Machine Learning Methods to Infer Precipitation Phase from Temperature and Moisture Profiles, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-16390,, 2021.

Nico Becker, Henning Rust, and Uwe Ulbrich

In Germany about 1000 severe road accidents are recorded by the police per day. On average, 8 % of these accidents are related to weather conditions, for example due to rain, snow or ice. In this study we compare several versions of a logistic regression models to predict hourly probabilities of such accidents in German administrative districts. We use radar, reanalysis and ensemble forecast data from the regional operational model of the German Meteorological Service DWD as well as police reports to train the model with different combinations of input datasets. By including weather information in the models, the percentage of correctly predicted accidents (hit rate) is increased from 30 % to 70 %, while keeping the percentage of wrongly predicted accidents (false-alarm rate) constant at 20 %. Accident probability increases nonlinearly with increasing precipitation. Given an hourly precipitation sum of 1 mm, accident probabilities are approximately 5 times larger at negative temperatures compared to positive temperatures. When using ensemble weather forecasts to predict accident probabilities for a leadtime of up to 21 h ahead, the decline in model performance is negligible. We suggest to provide impact-based warnings for road users, road maintenance, traffic management and rescue forces.

How to cite: Becker, N., Rust, H., and Ulbrich, U.: Translating weather forecasts to road accident probabilities, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-2092,, 2021.

Sinclair Chinyoka, Thierry Hedde, and Gert-Jan Steeneveld

Forecasting valley winds over complex terrain using a coarse horizontal resolution mesoscale model is a challenging task. Mesoscale models such as
the Weather Research and Forecasting (WRF) model tend to perform poorly over such regions. In this study, we assess the added value of downscaling
WRF wind forecasts using artificial neural networks (ANN) over the Cadarache Valley which is located in southeast France. Wind forecasts over the Cadarache valley are generated using WRF with a horizontal resolution of 3km on a daily basis. We used performance metrics such as Directional ACCuracy (DACC) and mean absolute error (MAE) for the evaluation of the WRF and ANN. WRF horizontal wind components at 110m and the near surface vertical potential temperature gradient were used as input data and observed horizontal wind components at 10m within the valley as targets during ANN training. We found an increase of DACC from 56% to 79% after post-processing WRF forecasts with ANN. Further analysis show that the ANN performed well during day and night, but poorly during morning and afternoon transition. The performance of WRF has a huge influence on ANN performance with bad WRF forecasts affecting ANN performance. However, the ANN improves the poor WRF forecasts to a DACC exceeding 60%. A change in lead time and domain resolution showed negligible impact suggesting that 3km resolution and a lead time of 24-47h is effective and relatively cheap to apply. Additionally, WRF performs well in near-neutral conditions and poorly in other atmospheric stability conditions. However ANN showed a consistent improvement in wind forecast during all stability classes with a DACC of nearly 80%. The study clearly demonstrates the ability to improve Cadarache valley wind forecasts using ANN from WRF simulations on a daily basis.

How to cite: Chinyoka, S., Hedde, T., and Steeneveld, G.-J.: Using an Artificial Neural Network to improve operational wind prediction in a small unresolved valley, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-3800,, 2021.

Christoph Fischer, Elmar Schömer, Andreas H. Fink, Michael Riemer, and Michael Maier-Gerber

Potential vorticity streamers (PVSs) are elongated quasi-horizontal filaments of stratospheric air in the upper troposphere related to, for example, Rossby wave breaking events. They are known to be related to partly extreme weather events in the midlatitudes and subtropics and can also be involved in (sub-)tropical cyclogenesis. While several algorithms have been developed to identify and track PVSs on planar isentropic surfaces, less is known about the evolution of these streamers in 3D, both climatologically but also for a better understanding of individual weather events. Furthermore, characteristics of their 3D shape have barely been considered as a predictor for high impact weather events like (sub-)tropical cyclones.

We introduce a novel algorithm for detection and identification of PVSs based on image processing techniques which can be applied to 2D and 3D gridded datasets. The potential vorticity was taken from high resolution isentropic analyses based on the ERA5 dataset. The algorithm uses the 2 PVU (Potential Vorticity Unit) threshold to identify and extract anomalies in the PV field using signed distance functions. This is accomplished by using a stereographic projection to eliminate singularities and keeping track of the reduced distortions by storing precomputed distance maps. This approach is computationally efficient and detects more interesting structures that exhibit the general behavior of PVSs compared to existing 2D techniques.

For each identified object a feature vector is computed, containing the individual characteristics of the streamers. In the 3D case, the algorithm looks at the structure en bloc instead of operating individually on multiple 2D levels. This also makes the identification stable regarding the seasonal cycle. Feature vectors contain parameters about quality, intensity and shape. In the case of 2D datasets, best-fitting ellipses computed from the statistical moments are regarded as a description of their shape. For 3D datasets, recent visualizations show that the boundary of these structures could be approximated by quadric surfaces . The feature vectors are also amended by tracking information, for example splitting and merging events. This low-dimensional representation serves as base for ERA5 climatologies. The data will be correlated with (sub-)tropical cyclone occurrence to spot useful and novel predictors for cyclone activity and preceding Rossby Wave Breaking events.

Overall, this new type of PVS identification algorithm, applicable in 2D or 3D, allows to diagnose the role of PVS in extreme weather events, including their predictability in ensemble forecasts.

How to cite: Fischer, C., Schömer, E., Fink, A. H., Riemer, M., and Maier-Gerber, M.: A novel identification and tracking method of weather-relevant 3D Potential vorticity streamers, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-9115,, 2021.

Andreas Beckert, Lea Eisenstein, Tim Hewson, George C. Craig, and Marc Rautenhaus

Atmospheric fronts, a widely used conceptual model in meteorology, describe sharp boundaries between two air masses of different thermal properties. In the mid-latitudes, these sharp boundaries are commonly associated with extratropical cyclones. The passage of a frontal system is accompanied by significant weather changes, and therefore fronts are of particular interest in weather forecasting. Over the past decades, several two-dimensional, horizontal feature detection methods to objectively identify atmospheric fronts in numerical weather prediction (NWP) data were proposed in the literature (e.g. Hewson, Met.Apps. 1998). In addition, recent research (Kern et al., IEEE Trans. Visual. Comput. Graphics, 2019) has shown the feasibility of detecting atmospheric fronts as three-dimensional surfaces representing the full 3D frontal structure. In our work, we build on the studies by Hewson (1998) and Kern et al. (2019) to make front detection usable for forecasting purposes in an interactive 3D visualization environment. We consider the following aspects: (a) As NWP models evolved in recent years to resolve atmospheric processes on scales far smaller than the scale of midlatitude-cyclone- fronts, we evaluate whether previously developed detection methods are still capable to detect fronts in current high-resolution NWP data. (b) We present integration of our implementation into the open-source “Met.3D” software ( and analyze two- and three-dimensional frontal structures in selected cases of European winter storms, comparing different models and model resolution. (c) The considered front detection methods rely on threshold parameters, which mostly refer to the magnitude of the thermal gradient within the adjacent frontal zone - the frontal strength. If the frontal strength exceeds the threshold, a so-called feature candidate is classified as a front, while others are discarded. If a single, fixed, threshold is used, unwanted “holes” can be observed in the detected fronts. Hence, we use transparency mapping with fuzzy thresholds to generate continuous frontal features. We pay particular attention to the adjustment of filter thresholds and evaluate the dependence of thresholds and resolution of the underlying data.

How to cite: Beckert, A., Eisenstein, L., Hewson, T., Craig, G. C., and Rautenhaus, M.: Objective 3D atmospheric front detection in high-resolution numerical weather prediction data, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-10055,, 2021.

Eigil Kaas and Emy Alerskans

Six adaptive post-processing methods for correcting systematic biases in forecasts of near-surface air temperatures, using local meteorological observations, are assessed and compared. The methods tested are based on the simple moving average and the more advanced Kalman filter - constructed to remove the longer-term bias, the very short-term errors or a combination of the two. Forecasts from a coarser-resolution global model and a regional high-resolution model are post-processed and the results are evaluated for one hundred private weather stations in Denmark. Overall, the postprocessing method for which a moving average is combined with a Kalman filter, constructed to remove the very short-term errors, performs the best. The biases of both the global coarserresolution forecasts and the regional high-resolution forecasts are reduced close to zero for all forecast lead times. The standard deviation is reduced for all forecast lead times for the coarser resolution model, whereas for the high-resolution model the most significant reduction is seen for the first six forecast lead hours. This shows that the application of a relatively simple postprocessing method, based on a short training period, can give good results.

How to cite: Kaas, E. and Alerskans, E.: Local temperature forecasts based on post-processing of Numerical Weather Prediction data, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-11270,, 2021.

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

Numerical weather prediction (NWP) models are known to exhibit systematic errors, especially for near-surface variables such as air temperature. This is partly due to deficiencies in the physical formulation of the model dynamics and the inability of these models to successfully handle sub-grid phenomena. Forecasts that better match the locally observed weather can be obtained by post-processing NWP model output using local meteorological observations. Here, we have implemented a non-linear post-processing model based on machine learning techniques with the aim of post-processing near-surface air temperature forecasts from a global coarse-resolution model in order to produce localized forecasts. The model is trained on observational from a network of private weather stations and forecast data from the global coarse-resolution NWP model. Independent data is used to assess the performance of the model and the results are compared with the performance of the raw NWP model output. Overall, the non-linear machine learning post-processing method reduces the bias and the standard deviation compared to the raw NWP forecast and produces a forecast that better match the locally observed weather.

How to cite: Alerskans, E., Nyborg, J., Birk, M., and Kaas, E.: Prediction of near-surface temperatures using a non-linear machine learning post-processing model, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-11378,, 2021.

Natacha Galmiche, Nello Blaser, Morten Brun, Helwig Hauser, Thomas Spengler, and Clemens Spensberger

Probability distributions based on ensemble forecasts are commonly used to assess uncertainty in weather prediction. However, interpreting these distributions is not trivial, especially in the case of multimodality with distinct likely outcomes. The conventional summary employs mean and standard deviation across ensemble members, which works well for unimodal, Gaussian-like distributions. In the case of multimodality this misleads, discarding crucial information. 

We aim at combining previously developed clustering algorithms in machine learning and topological data analysis to extract useful information such as the number of clusters in an ensemble. Given the chaotic behaviour of the atmosphere, machine learning techniques can provide relevant results even if no, or very little, a priori information about the data is available. In addition, topological methods that analyse the shape of the data can make results explainable.

Given an ensemble of univariate time series, a graph is generated whose edges and vertices represent clusters of members, including additional information for each cluster such as the members belonging to them, their uncertainty, and their relevance according to the graph. In the case of multimodality, this approach provides relevant and quantitative information beyond the commonly used mean and standard deviation approach that helps to further characterise the predictability.

How to cite: Galmiche, N., Blaser, N., Brun, M., Hauser, H., Spengler, T., and Spensberger, C.: Exploring multi-modalities in weather prediction using a univariate graph based on machine learning techniques, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-11747,, 2021.

Samantha Ferrett, Thomas Frame, John Methven, Christopher Holloway, Stuart Webster, Thorwald Stein, and Carlo Cafaro

Forecasting extreme rainfall in the tropics is a major challenge for numerical weather prediction. Convection-permitting (CP) models are intended to enable forecasts of high-impact weather events. Development and operation of these models in the tropics has only just been realised. This study describes and evaluates recently developed Met Office Unified Model CP ensemble forecasts of varying resolutions over three domains in Southeast Asia, covering Malaysia, Indonesia and the Philippines.

Fractions Skill Score is used to assess the spatial scale-dependence of skill in forecasts of precipitation during October 2018 - March 2019. CP forecasts are skilful for 3-hour precipitation accumulations at spatial scales greater than 200 km in all domains during the first day of forecasts but all ensembles have low spread relative to forecast skill. Skill decreases with lead time and is highly dependent on the diurnal cycle over Malaysia and Indonesia. Skill is largest during daytime when precipitation is over land and is constrained by orography, but is lower at night when precipitation is over the ocean. Comparisons of CP ensembles using 2.2, 4.5 and 8.8 km grid spacing and an 8.8km ensemble with parameterised convection are made to examine the role of resolution and convection parameterisation on forecast skill for the three domains.

How to cite: Ferrett, S., Frame, T., Methven, J., Holloway, C., Webster, S., Stein, T., and Cafaro, C.: Evaluating convection-permitting ensemble forecasts of precipitation over Southeast Asia , EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-12375,, 2021.

Carlo Cafaro, Beth J. Woodhams, Thorwald H. M. Stein, Cathryn E. Birch, Stuart Webster, Caroline L. Bain, Andrew Hartley, Samantha Clarke, Samantha Ferrett, and Peter Hill

Convection-permitting ensemble prediction systems (CP-ENS) have been implemented in the
mid-latitudes for weather forecasting timescales over the past decade, enabled by the increase in
computational resources. Recently, efforts are being made to study the benefits of CP-ENS for
tropical regions. This study examines CP-ENS forecasts produced by the UK Met Office over
tropical East Africa, for 24 cases in the period April-May 2019. The CP-ENS, an ensemble with
parametrized convection (Glob-ENS), and their deterministic counterparts are evaluated against
rainfall estimates derived from satellite observations (GPM-IMERG). The CP configurations have
the best representation of the diurnal cycle, although heavy rainfall amounts are overestimated
compared to observations. Pairwise comparisons between the different configurations reveal that
the CP-ENS is generally the most skilful forecast for both 3-h and 24-h accumulations of heavy
rainfall (97th percentile), followed by the CP deterministic forecast. More precisely, probabilistic
forecasts of heavy rainfall, verified using a neighbourhood approach, show that the CP-ENS is
skilful at scales greater than 100 km, significantly better than the Glob-ENS, although not as good
as found in the mid-latitudes. Skill decreases with lead time and varies diurnally, especially for
CP forecasts. The CP-ENS is under-spread both in terms of forecasting the locations of heavy
rainfall and in terms of domain-averaged rainfall. This study demonstrates potential benefits in
using CP-ENS for operational forecasting of heavy rainfall over tropical Africa and gives specific
suggestions for further research and development, including probabilistic forecast guidance.

How to cite: Cafaro, C., Woodhams, B. J., Stein, T. H. M., Birch, C. E., Webster, S., Bain, C. L., Hartley, A., Clarke, S., Ferrett, S., and Hill, P.: Do convection-permitting ensembles lead to more skilful short-range probabilistic rainfall forecasts over tropical East Africa ?, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-13689,, 2021.

Kameswarrao Modali and Marc Rautenhaus

Ensemble forecasting has become a standard practice in numerical weather prediction in forecasting centres across the world. The large data sets generated by ensemble forecasting systems carry much information, that is difficult to analyse in short time periods, requiring well-designed workflows in order to be useful.

Clustering is one of the ensemble analysis methods that are applied to discover similarities between ensemble members. Cluster analysis involves different steps like dimensionality reduction, core clustering algorithm and evaluation. A large of number of methods have been proposed in the literature for each of these steps, however, only few have been applied to clustering of ensemble forecasts. A major challenge is that for a given ensemble forecast, different choices of methods and data domains can lead to very different clustering results. For example, Kumpf et al. (2018, IEEE Transact. Vis. Comp. Graph.) have demonstrated the sensitivity of clustering results to even small changes in the considered domain. The challenge equally exists for choices in clustering methods and method parameters.

In our work, we are attempting to open up the clustering black box by introducing a visualization workflow that makes transparent to the user how different choices in methods and method parameters lead to different clustering results. To achieve this, a clustering analysis library that works in tandem with the ensemble visualization software “Met.3D” () is being developed. We present the current state of the system and demonstrate its use by analysing an ensemble forecast case study.

How to cite: Modali, K. and Rautenhaus, M.: Bringing transparency into ensemble cluster analysis with the aid of interactive visualization, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-12854,, 2021.

Babitha George and Govindan Kutty

Ensemble forecasts have proven useful for investigating the dynamics in a wide variety of atmospheric systems and they might be useful for diagnosing the source of forecast uncertainty in multi-scale flows. Ensemble Sensitivity Analysis (ESA) uses ensemble forecasts to evaluate the impact of changes in initial conditions on subsequent forecasts. ESA leads to a simple univariate regression by approximating the analysis covariance matrix with the corresponding diagonal matrix. On the contrary, the multivariate ensemble sensitivity computes sensitivity based on a more general multivariate regression that retains the full covariance matrix. The purpose of this study is to examine the performance of multivariate ensemble sensitivity over univariate by applying it to a heavy rainfall event that happened over the Himalayan foothills in June 2013. The ensemble forecasts and analyses are generated using the Advanced Research version of the Weather Research and Forecasting (WRF) model DART based Ensemble Kalman Filter. Initial results are promising and the sensitivity shows similar patterns for both univariate and multivariate methods. The reflectivity forecast for both methods are characterized by lower temperatures and increased moisture in the control area at 850 hPa level. Compared to multivariate, univariate ensemble sensitivity overestimates the magnitude of sensitivity for temperature. But the sensitivity for the moisture is the same in both methods.

How to cite: George, B. and Kutty, G.: Understanding the Dynamics of a Heavy Rainfall Event using Multivariate Ensemble Sensitivity Analysis , EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-11954,, 2021.