Evaluating the Effect of Additional Data Channels for CNN-based Radar Precipitation Nowcasting

Peter Pavlík

PhD student at slovak.AI

Abstract: Nowcasting in meteorology is defined as forecasting with high local detail, by any method, over a period from the present to six hours ahead. In general, the main nowcasting task is to predict precipitation amounts in the near future and generate alerts on extreme weather events, preventing damage and saving lives during flash floods or hail storms. In the past years, data-driven approaches to precipitation nowcasting have started to gain attention, attempting to mitigate the main limitations of the traditional optical flow approaches by using deep neural networks to learn and predict the dynamic patterns in radar reflectivity precipitation observations. Deep learning models can learn complex patterns while still satisfying the requirement of producing an output quickly. We propose a comprehensive investigation into the impact of additional data channels on radar precipitation nowcasting using convolutional neural networks. The primary objective is to evaluate the effectiveness of integrating various additional weather variables into the nowcasting model to provide the model with enhanced situational awareness and a deeper understanding of atmospheric processes. The study will leverage the Royal Netherlands Meteorological Institute Data Platform to obtain radar reflectivity data and assimilate it with selected additional datasets. The research will be conducted over a three-month period and aims to produce a research paper summarizing the findings, an open-source codebase for reproducibility, and a trained nowcasting model that we believe can outperform the baseline model in terms of accuracy and reliability.

Keywords: convolutional neural networks, earth observation, remote sensing, precipitation nowcasting
Scientific area: Artificial Intelligence – Machine Learning in Geosciences
Visiting period: 08/01/2024 – 05/04/2024
Visiting Lab: Delft University of Technology (Dept. of Geoscience and Remote Sensing)

Bio: Peter is a third year student at the Kempelen Institute of Intelligent Technologies in Bratislava (a member of slovak.AI). He uses deep learning to improve the efficiency and accuracy of forecasts of physical systems. His dissertation thesis focuses on leveraging physics-informed machine learning in forecasting, mainly precipitation nowcasts.