The current project is aimed to bring together the expertise from three research and application domains: the domain of intelligent systems and methods, the field of software development and engineering and the operative and research in the field of meteorology. According to the World Meteorological Organisation, many natural disasters that cause damage and loss of life are due to weather, and especially severe weather. Most of the meteorological decision-making processes rely to a great extent, at least at the national level, on the experience of operative meteorologists, but due to the chaotic character of the atmosphere, effectively and accurately analyzing a large volume of meteorological data is a difficult task for meteorologists. WinDMiL project aims to contribute to improving meteorological decision support at the national level in terms of faster decision times and higher precision, providing a working prototype of a software solution for remote sensing data mining and assisting decision-making processes in meteorology. The direct beneficiary of WinDMiL is the Romanian National Meteorological Administration. WinDMiL will integrate data mining methods for fast and precise remote sensing data analysis, machine learning methods for weather forecast and a component for visualizing the results of the mining, easily readable for meteorologists and aimed to facilitate meteorological decision-making.
Project Outcome
The current project proposal aims to provide a working prototype of a ML software solution based on mining remote sensing data for assisting operational meteorologists and meteorological researchers in weather forecasting, thus contributing to improving meteorological decision support at the national level in terms of faster decision times and higher precision. The main result of the project will be WinDMiL, a new DM software solution for analyzing remote sensing data and assisting decision-making processes in meteorology. The direct beneficiary of WinDMiL is the Romanian NMA, but all meteorological institutes (meteorologists and meteorological researchers) may benefit from the project result. WinDMiL will integrate DM methods for fast and precise data analysis and will include a component for visualizing the results of the mining, easily readable for meteorologists and aimed to facilitate meteorological decision-making, mainly regarding the occurrence of severe weather phenomena (rainfall, hail, etc.). The goal of the DM component is twofold: (1) to analyse past/historical remote sensing data (radar, satellite data) to uncover patterns regarding the weather evolution; and (2) prediction of weather maps and parameters through forecasting of radar and satellite products. Remote sensing data are commonly used by meteorologists to analyse and predict the evolution of weather. Apart from existing solutions, we aim to use fused radar and satellite data which is likely to improve the performance of forecasting. While this can be hard for reasons such as the difference between scales and data types, fusing satellite and radar data should improve the DM models as they represent different aspects of the same weather phenomenon. By employing ML techniques, the project intends to broaden the use of remote sensing data in the direction of reanalysis of the historical data. The results of such analyses would be relevant in various meteorological processes such as quantitative precipitation forecast and weather maps prediction. These are among the most challenging activities in meteorology and are useful in forecasting severe events such as hail, rainfall, etc. The DM software solution proposed in the current project may provide meteorologists and climatologists the possibility to analyze a huge database containing relevant information on weather evolution, which otherwise would be very difficult to do in a formal manner.
Objectives
O1. Development and scientific validation of new DM techniques specially designed for remote sensing data analysis, with a specific emphasis on radar and sattelite data fusion.
Following our previous scientific results in the UL based analysis of meteorological data as in DL-based weather nowcasting, methods such as deep k-nearest neighbor (kNN), deep CBR and deep convolutional AEs and deep recurrent encoder-decoder architectures specially tailored for remote sensing data will be targeted as DL models for components C1 and C2.
O2. Development and user evaluation of the WinDMiL system for descriptive and predictive DM for graphical visualization of the mining results.
This will be achieved using the methods developed as part of O1 and is expected to support meteorological decision-making.
Project financed under PN-IV-P7-7_1-PED-2024-0121, contract nr. 13PED/2025