WeaMyL project aimed at enhancing the accuracy, performance and reliability of national nowcasting warning systems by the use of machine learning (ML) techniques applied on radar, satellite and weather stations' observations. The focus was on obtaining higher precision in predicting the occurrence and the areas affected by severe meteorological phenomena, as well as attaining lower decision times (compared to former, exclusively human decision times). The project's main goal was to automate the nowcasting warning systems by creating a ML driven platform for early and accurate forecast of severe phenomena. Thus, it was aimed to be the backbone of a new framework for imminent severe weather detection adapted to current technological possibilities.
Main objectives
The main objectives of the project were the following:
Development and scientific validation of novel ML computational models and techniques specially tailored for accurate nowcasting.
-
Development and user evaluation of the Annotated Atlas of Meteorological Observations, a large database containing meteorological data (radar, satellite and other relevant meteorological observations) being used as a data source for the Deep Learning techniques developed within the project (data extraction, processing and classification).
Development of the open-source WeaMyL platform for early forecast of severe weather phenomena. The objective was to provide a big-data based forecasting platform that includes the Supervised Learning methods developed and which supports meteorologists in improving the quality of nowcasting alerts.
-
Contribute to the development of scientific knowledge by disseminating the obtained scientific results through scientific publications, the project website and social media.
WeaMyL software platform
The WeaMyL software platform consists of three components:
-
Forecasting Platform. Given that the accuracy of the prediction module is paramount for the system's precision, this is WeaMyL's central component. The forecasting platform employs historical and real-time data and is responsible with estimating the location, time and severity of meteorological phenomena expected to take place in the near future (0-6 hours). Our algorithms model the meteorological situation based on the earlier meteorological situations. The employed learning models find patterns in the meteorological data which apply to specific circumstances and thus improve themselves as increasingly more data becomes available, hence making the forecasting module self-adaptive. To process the available amount of meteorological data and to accelerate scientific analysis and information extraction, big data approaches were targeted.
-
Annotated Atlas of Meteorological Observations The Annotated Atlas component provides a semantic database over the large volume of historical data from varying sources, including weather stations' observations, radar and satellite imagery. Its main objectives are to facilitate the study of past conditions using statistical and comparative analyses, provide intelligent information retrieval based on a semantic data model, and provide a custom, extensible annotation model across observation types and sources. The Atlas component offers two main functionalities: (a) management of the Atlas and (b) visualizations, statistics and reports based on the historical meteorological data stored in the Atlas.
-
Integration Module. In order for the above components to provide actionable information, they must be integrated with the software systems already deployed on location. The purpose of this module is to provide a number of data consumer and data provider connectors that can be integrated within existing systems. Both the Forecasting Platform and the Annotated Atlas act as consumers for real-time generated meteorological data, while the Forecasting Platform provides data relevant to upcoming severe weather.
Work Packages
The work programme was divided into five work packages (WPs), following the usual stages needed to reach the main research objective, namely to develop the WeaMyL software platform. These are:
WP1 - Documentation, system requirements and architecture
The first work package consisted on documentation, study, literature analyses and identifying the limitations of existing state-of-the-art approaches and solutions in nowcasting. Afterwards, the functional and nonfunctional requirements for WeaMyL including the end-user requirements, as well as the main conditions for the system's proper functioning were defined, in order to establish the general architecture and design of the WeaMyL software platform.
WP2 - Machine learning models for weather nowcasting
The second work package consisted of defining a theoretical model for nowcasting along with (2) Develop specially tailored scalable ML models for accurate nowcasting. (3) Scientifically validate developed ML models using experimental results analysis and interpretation, as well as comparison to related work.
WP3 - Software development, testing and integration
WP3 was focused on software development, testing, and platform integration with national warning systems. The Forecasting Platform is comprised of the ML components (accessible via an API) together with the front-end. An effort was made concerning the integration of WeaMyL with national warning systems. The platform prototype was developed in an incremental, iterative manner, having as final objective its successful integration with Romanian and Norwegian weather warning systems.
WP4 - Meteorological evaluation, interpretation and analysis
WP4 activities included extracting, annotating and validating relevant meteorological data from NMA and MET databases and piloting the platform within NMA and MET locations. These activities were coordinated by NMA in close cooperation with MET-MT. Results concerning the accuracy, performance and reliability of the platform were discussed with MET-IT and BBU. The feedback was used to continuously improve the ML modules, front-end components and the Annotated Atlas.