Opening Ceremony
| [09:50 – 10:00] Opening the Workshop To be announced |
Meteorological Research
[10:00 – 10:30] A country scale assessment of the heat hazard-risk in urban areas
Sorin Cheval (National Meteorological Administration, Romania)
The connection between the regional climate and urban environment produces local changes in most climate features and exacerbates the magnitude and frequency of extreme events. Heat hazards are particularly related to urban climate and global warming will amplify the associated risks in vulnerable areas and their environmental impacts. In this talk we present the results of interdisciplinary environmental research aiming to provide a country scale perspective of the heat hazard-risk (HHR). It is used as a risk matrix approach combining elements of thermal hazard, derived from land surface temperature (LST), and vulnerability metrics, derived from population density and urban fabric. The study informs about the overall factors that control the HHR across cities in Romania, such as the environmental and climate settings, the city size and structure. The results show that the urban HHR is higher during the daytime, in warmer climates and in densely populated cities. The use of the methodology at a country scale is innovative, and demonstrates clear potential for applications in other regions, mainly for national strategies and plans aiming to mitigate the urban HHR within the climate change context.
[10:30 – 11:00] Variability of precipitation in Romania: recent trends and projected changes in climate scenarios
Mihaela Caian (National Meteorological Administration, Romania)
We analyse precipitation trends and variability and their mechanism during the past decades over Romania. Main changes e.g. a positive trend towards autumn (September – October, 1961-2016) in extra-Carpathian regions and spring (during March – April) and negative in summer are shown to be linked to changes in the large scale drivers: the polar jet and the main modes of atmospheric variability and their interaction with surface climatological sources of latent heat as Mediterranean and Black Seas cyclogenesis. These impact our regions by changing the seasonal frequency and intensity of storm tracks crossing the region. The projected variability under climate change is also analysed in climate scenarios. The driving mechanism for regional changes remains valid as indicated by the storm tracks computed for climate scenarios. Under a warmer climate, the high resolution signal from CORDEX/CMIP5 appears regionally well correlated with the newest simulations in CMIP6, with two notices: one is that CMIP6 enhances even more the already estimated signal in CMIP5 (for both precipitation and temperature) and the other is related to further impact applications, as long as these show a non-linear response to the combined socio-economic / climate scenarios that should be carefully accounted for at regional scale.
[11:00 – 11:30] How nowcasting is used in today’s danger warnings
Magnus Haukeland (Norwegian Meteorological Institute)
We will give an overview of how nowcasting is used in today’s operational forecasting, focusing on danger warnings. Nowcasting is mainly used for issuing/updating warnings for rain flood and lightning, so these will be the main topics. We will present a few case studies where nowcasting was an essential tool used to issue danger warnings, and to consider the danger level of the warnings.
[11:30 – 12:00] Processing and filtering of radar end-product composite reflectivity
Christoffer Artturi Elo (Norwegian Meteorological Institute)
This presentation will give a short summary on all the methods that are currently in use for giving clean and visually pleasant radar images for presentation on yr.no. Radar images are cleaned by identifying various clutter such as ground – and sea clutter, sun flare, RLAN, ship traffic and other spurious echoes. The ground clutter is important to not only identify, but to remove and replace with surrounding data, this greatly enhances nowcasting algorithms since the clutter is stationary in time.
WeaMyL Research
[12:30 – 13:00] Using convolutional autoencoders for precipitation nowcasting based on radar data
Mihai Andrei (Babeș-Bolyai University)
Weather radar data is of utmost importance to meteorologists for issuing short-term weather forecasts and warnings of severe weather phenomena. During our research we also experimented with precipitation prediction in terms of classification: predict whether the radar reflectivity values will be above or below a certain threshold in the next time step. We proposed a machine learning binary classification model based on two convolutional autoencoders intended for precipitation nowcasting by predicting weather radar reflectivity. The autoencoders were trained on radar data collected on both stratiform and convective weather conditions. AutoNowP is intended to be a proof of concept that autoencoders are useful in distinguishing between convective and stratiform precipitation. Real radar data provided by the Romanian National Meteorological Administration and the Norwegian Meteorological Institute is used for evaluating the effectiveness of AutoNowP. Results showed that AutoNowP surpassed other binary classifiers used in the supervised learning literature in terms of probability of detection and negative predictive value, highlighting its predictive performance.
[13:00 – 13:30] ConvLSTM architectures for meteorological nowcasting based on satellite imagery
Mircea-Ioan Gabriel (Babeș-Bolyai University)
The satellite imagery repositories in the custody of MET offer one of the richest collections of meteorological data in Europe, therefore capitalizing on this informational golden ore was only natural in the context of our purposes of developing deep learning models for meteorological nowcasting. As it promised to be a model that can pick up on dynamics similar to the motion of meteorological fronts in the atmosphere, the ConvLSTM model was chosen as the basis for several deep learning architectures employed with the task of predicting meteorological products with respect to prior consecutive temporal satellite readings. The trained models show the promise of adaptation to the temporal progression of meteorological products, granted the model is trained over a considerable number of iterations on significantly computationally demanding data sources.
[13:30 – 14:00] How to use the WeaMyL machine learning forecasting platform
Abdelkader Mezghani (Norwegian Meteorological Institute)
We will give a quick demo on how to use the machine learning forecasting platform from an IT perspective. The platform combines a crontab job to automatically generate a forecast issued every 5 minutes, a THREDDS data server to deal with data I/O and web mapping tools to visualize the data. We will present a short live demo of these components and create animations of forecasted versus observed data such as Reflectivity data using the MET Norway meteorological Radar Network. We will also introduce a few web mapping tools such as Godiva2 and OpenGeoWeb that are fully compatible with MET Norway data I/O located on the MET TDS.
[14:00 – 14:30] A brief introduction on the netCDF file format, the OPeNDAP protocol and THREDDS Data server
Arild Burud (Norwegian Meteorological Institute)
In this presentation we will show how netCDF files are used as a medium for meteorology data exchange, along with examples on how to use the THREDDS Data Server and OPeNDAP protocol for data extraction. Finally we will give a few notes on the installation of a local THREDDS server.
Applications of Deep Learning
[15:00 – 15:30] MBMT-Net: A Multi-Backbone, Multi-Task Deep Convolutional Neural Network
Ciubotariu George (Babeș-Bolyai University)
Recently proposed improvements in the field of Computer Vision (CV) refer to either enhancing the feature extraction capabilities of encoders or the study of convolutional building blocks by adding more parameters or using advanced techniques and mechanisms, which ultimately aid Single-Task (ST) Convolutional Neural Networks (CNN) achieve state-of-the-art results in solving dense tasks. However, the problem formulations may be suboptimal in what regards the learning context they are put in. Consequently, Multi-Task (MT) learning has been introduced to jointly train multiple tasks at once, which empowers MT architectures with more consistent and robust feature extraction means. Nevertheless, MT-CNNs may lack the capacity to extract various features, since by sharing the backbone all the gradients are propagated into it, possibly resulting in its oversaturation; this must be overcome by using deeper encoders or Multi-Backbone (MB) feature extractors. Hence, as a strategy proposed to compensate for these shortcomings, we introduce MBMT-Net, a MB-MT-CNN architecture based on a development strategy that outperforms both ST and MT approaches by infusing backbones with more diverse and specialised processing capabilities. Due to using pre-trained backbones for particular tasks, the concatenated feature maps offered to the MT heads are guaranteed to have richer and more specific features with decreased network parameters when compared to traditional MT architectures. Our strategy is architecture independent, and it can be applied to different types of backbones and parsing heads, which greatly extends the domain of configurable features, finally providing us with great experimental support for MB and MT learning.
[15:30 – 16:00] Generation and Simulation of Artificial Human Societies using Anthropologically modelled Learning Agents
Galan Radu (Babeș-Bolyai University)
As soon as we will comprehend the inner mechanisms of human societies, then we will also be able to answer more pragmatical and existential questions about humans. Simplified models allow us to understand these sophisticated concepts and identify basic correlations. Artificial society models will most likely explain and prove these correlations between “micro-level cognition and macro-level social behavior”. Using a custom agent-based system we can provide a scalable architecture that can be the foundation for complex social simulations. The environmental attributes (temperature, humidity, altitude, vegetation spread, fauna density etc.) will need to be either iterated using available historical data or modelled based on known parameters and simulated. Each agent is going to perceive the environment, conduct an internal process of cognition and eventually act upon the world. All the processes surrounding the environment and the agents must be modelled as accurately as possible (in the computational limits available) based on anthropological, biological, geographical, botanical, and zoological statistical knowledge. The social simulation field currently focuses more on specific human situations (evacuation of buildings or airplanes, crowd control and traffic flow optimization) and less on general simulated society. This factor of originality must also find statistical validation through enough experimentation and proper confirmations of every correlation. The first experiment will aim to prove the correlation between thriving tribes and altruism as stated by the group selection theory.
[16:00 – 16:30] Application of Machine Learning algorithms for the botanical origin assessment of honey using isotope and elemental profiles
Hațegan Ariana-Raluca (Babeș-Bolyai University)
Honey adulteration detection represents an increasing preoccupation among researchers and control laboratories, a fact that has illustrated the need for trustworthy data processing tools able to unequivocally determine the authenticity of honey samples. In this context, we imagine the use of Machine Learning in order to construct honey differentiation models to be a step forward. In the present talk, we illustrate the applicability of Artificial Neural Networks (ANN) and Support Vector Machines (SVM), for classifying honey with respect to its botanical origin. For this purpose, the isotope and elemental profiles corresponding to an authentic honey sample set, having as botanical source acacia, colza, honeydew, linden and sunflower, were used as input data. A special attention was given to the pre-processing phase in order to identify the attributes that have a higher classification power. The performance parameters of the developed models indicate that ANN and SVM can be applied as reliable tools for assessing the botanical origin of honey.
[16:30 – 17:00] Protein-Protein Interaction Prediction using a Deep Learning Approach based on Autoencoders
Albu Alexandra (Babeș-Bolyai University)
Reliable detection of interactions between proteins represents an important goal in bioinformatics, which can give insights into molecular mechanisms, as well as provide a valuable tool for drug design. We present a deep learning sequence-based approach for detecting protein-protein interactions (PPIs) using a pair of autoencoders. The approach consists of training a separate autoencoder on each of the two classes – interacting and non-interacting – with the aim of learning high-level features which are characteristic for the pairs in the corresponding class. In order to better capture information from paired inputs, we design two siamese autoencoder architectures and compare them with a baseline architecture. We show that our approach surpasses several machine learning algorithms proposed in the literature on the considered data sets. In addition, we discuss limitations of current sequences-based PPI prediction approaches, alongside directions of future research.