Séminaire le 17 jan 2019 à 14h00
Salle Coriolis Salle Coriolis Observatoire Midi-Pyrénées - 14, avenue Edouard Belin - 31400 Toulouse

Voir le plan d'accès

Exploring Deep Learning Algorithms in Forecasting the Occurrence of Severe Haze Events in Southeast Asia

Adverse environmental and weather events cause substantial economic loss. For instance, particulate pollution has become a serious environmental and societal issue of many Southeast Asian countries in recent decades. Forecasting the occurrence of these high-impact events could allow mitigation measures to be implemented ahead of time and thus effectively minimize economic loss. In this front, the “traditional” numerical forecasting models seem to be the first choice. Nevertheless, due to the lack of real-time data such as emissions (e.g., for haze specifically), significant and sophisticated spatiotemporal variabilities associated with these often low-probability events, and the propagation of numerical or parameterization errors through model integration (well, ensemble could be a much expensive mitigation), such a task still poses a serious challenge to these “traditional” forecasting models. On the other hand, in the last few years deep machine learning has earned explosively growing popularity by conquering one after another changing task from visual and voice recognition, natural language, to game playing. Differing from many “traditional” machine learning techniques that often proceed following certain abstract models or procedures made by human, deep learning directly connects raw data as input with the output through deep artificial neural networks that decompose then learn different characters of the sophisticated structure of input data (e.g., images), providing an “end-to-end” solution. The application of deep learning in Earth sciences seem still in its earliest stage. Arguably the first group of published such attempts in identifying several weather patterns including tropical cyclone has been successful. The capability of deep neural networks in capturing very complex images indeed suggests a great potential for wider applications in our sciences. Here I will describe an exploratory effort of using deep CNN to forecast the occurrence of an adverse environmental event, i.e., haze or severe particulate pollution in Southeast Asia.

Afficher le pied de page