26–27 Feb 2020
RMIB
Europe/Brussels timezone

Short-term rainfall analysis by spectral clustering in Ivory Coast

Not scheduled
20m
Conference Room (RMIB)

Conference Room

RMIB

Ringlaan 3 B-1180 Brussels Belgium

Speaker

Mrs kelly Grassi (Ifremer, UlLCO LISIC, WeatherForce)

Description

Over the last decade, extreme events (i.e., storms, floods, droughts) are more and more frequent and/or intensive. 85% of the world's population is affected by these events. 70% of the world economy is dependent on meteorology according to the United Nations. 75% of the most vulnerable countries has no or little reliable, accurate and effective weather information. Indeed, some National Meteorological Services (NMS) have difficulties to deploy and maintain operational infrastructure like rain gauge recorder. In addition, rain gauges provide only local information, measuring rainfall level in the specific geographic location. Furthermore, it is more complex to efficiently forecast extreme events, such as heavy rain, the onset of monsoons, in specific area. In this case only atmospheric model data, like climate reanalysis (ERA5), are fully available.

The Ivory Coast NMS (Sodexam) are interested in interpreting and forecasting rainfall events based on ERA5 time series. It is a fundamental information to preserve off-ground culture, infrastructure or local agricultural production such as cocoa. Sodexam defined rainfall criterion based on 4 intensities: no rain, weak rain, moderate rain, and heavy rain. Based on this criterion, many supervised classification methods have been compared. However, no method obtained satisfying results due to large confusion between light and moderate rainfalls that occur in same environmental conditions. Heavy rain is also difficult to identify because
this cluster is under-represented in the dataset.

To understand the dynamics of rainfall events, unsupervised classification approach is an alternative solution. It allows the extraction of pattern guided by data geometry. In this study, we apply unsupervised clustering method, Multi-level Spectral Clustering (MSC), to extract patterns characteristic, independent of rainfall thresholds. MSC already succeeded in detecting environmental patterns in marine multivariate time series (Grassi et al., 2019) or georeferenced dataset (DYPHYMA campaign; Lefebvre et Caillault, 2019). MSC is based on divisive clustering conducted in eigenspace from data similarity. MSC provide a deep representation from general environmental patterns to extreme events. We conduct same methodology on ERA5 to detect temporally structured successive events. Dynamics of the obtained clusters/events are coherent with monsoon and rainy season.

References:

  • K. GRASSI, E. POISSON CAILLAULT and A. LEFEBVRE, "Multilevel Spectral Clustering for extreme event characterization," OCEANS 2019 - Marseille, Marseille, France, 2019, pp. 1-7. doi: 10.1109/OCEANSE.2019.8867261 URL: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8867261&isnumber=8866872

  • DYPHYMA campaign: 2012 Continuous Phytoplankton measurements (DYPHYMA), springer 2012, eastern channel. https://sextant.ifremer.fr/record/5dbafe69-81cf-4202-b541-9f8b564fa6f9/

  • A. Lefebvre, Emilie Poisson Caillault.(2019) High resolution overview of phytoplankton spectral groups and hydrological conditions in the eastern English Channel using unsupervised clustering. Marine Ecology Progress Series, Inter Research, 2019, 608, pp.73-92.doi:10.3354/meps12781.

Primary author

Mrs kelly Grassi (Ifremer, UlLCO LISIC, WeatherForce)

Co-authors

Presentation materials

There are no materials yet.