Rain nowcasting with Radar reflectivity maps

Encadrants : Dominique Béréziat
Disponible : OUI

Spécialité : IMA
Nombre d'étudiants : 2
Description : Rainfall forecasting is an important environmental issue (flood anticipation, agriculture management...). Here your are interested in a short-time forecast (time horizon ranges from 30 minutes up to 2 hours). We use radar reflectivity maps which are robust to rainfall indicators. The literature of rain nowcasting using DL approches is quite recent (2015). From this date up to 2021, the problem has been formulated as classification task with less than 10 precipitation classes, such as no rain, light rain, moderate rain, etc.
Pré-requis : pytorch programming.
Travail demandé : In this work, we want to address the rain nowcasting as a regression problem, and predict rain precipitation as a physical quantity and not as a class. This made the problem harder. To help, students are invited to draw inspiration from two recent papers [1,2].
Liens complémentaires : [1] ED-DRAP: Encoder–Decoder Deep Residual Attention Prediction Network for Radar Echoes. Che et al. Geoscience and remote sensing letters. vol 19, 2022. [2] SmaAt-UNet: Precipitation Nowcasting using a Small Attention-UNet Architecture. Trebing et al. PRL. vol 145, 2021.