Deep learning for quality control of crowdsourced data and seamless weather forecasting. (OZR3893)

The amount and diversity of weather observations, such as
crowdsourced and novel satellite observations, has increased
tremendously in recent years. Together with high-resolution weather
models, these new data sources have the potential to yield better
forecasts and protect society from climate extremes. However, current
weather forecasting systems do not make full use of this vast amount
of heterogeneous and often noisy data. To make optimal use of the
increasing volume of crowdsourced observations, we will develop a
deep learning-based emulator to perform context-aware quality
control.
A second challenge is how to combine nowcasts (short-term forecasts
based on the extrapolation of observations) and different weather
models to obtain a single seamless and accurate weather forecast.
In this research project we will use deep learning (DL) to address these
challenges. Deep learning is a very promising branch of machine
learning, a subfield of artificial intelligence in which a computer
program learns from a set of training data. Deep learning uses neural
networks with many layers to learn abstractions and capture highly
complex relationships, such as the physical laws that drive atmospheric
motions. We will develop new domain-specific DL architectures that
learn how to optimally combine information from multiple weather
Research Council Regulations | Chapter 5, Article 19 – Centraal Werkingsreglement Onderzoek | approved AB 17.02.2020
Research proposal description clearly add expected outcome; use up to 1500 words
Motivation
Extreme weather events have a tremendous influence on our society and economy, as the
devastating floods of last summer made painfully clear. In addition to the 41 lives lost, the
economic cost of this extreme event will surpass € 2 billion in Belgium alone. Climate change
is worsening many severe weather events, such as extreme precipitation and heat waves
(IPCC, 2021). Accurate observations, forecasts and warnings are therefore crucial to prevent
deaths, injuries or damage caused by severe weather.
Recent years have seen a steady increase in the resolution and availability of multisensory
meteorological observations, including satellite and heterogeneous crowdsourced data. At
the same time, great progress has been made in improving high-resolution physics-based
numerical weather prediction (NWP) models. The progress of one additional day of forecast
skill per decade is as much due to improved modelling as due to increased availability of data
and improved data assimilation techniques (Bauer et al., 2015).
models and observations, and that can correct systematic errors
present in these models.

Keywords:Deep learning, crowdsourced observations, personal weather stations, climate adaptation, urban climate, seamless forecasting, nowcasting, data assimilation

Disciplines:Machine learning and decision making, Meteorology, Urban physics, Pattern recognition and neural networks