Exploiting deep learning in forecasting the occurrence of severe haze in Southeast Asia

TL;DR

This paper presents a deep learning framework using convolutional neural networks to forecast severe haze events in Southeast Asia, trained on 35 years of meteorological data, showing promising accuracy improvements over baseline methods.

Contribution

The study develops a novel deep learning-based forecasting model for severe haze events, demonstrating its potential in predicting extreme environmental phenomena.

Findings

Achieved high overall accuracy in haze prediction

Outperformed no-skill baseline forecasts

Identified challenges with false negatives

Abstract

Severe haze or low visibility event caused by particulate pollution has become a serious environmental issue in Southeast Asia. A forecasting framework of such events based on deep convolutional neural networks has been developed. The framework has been trained using time sequential maps of up to 18 meteorological and hydrological variables alongside surface visibility data over past 35 years. In forecasting haze versus no-haze situations in Singapore, the trained machine has achieved a good overall accuracy that easily exceeds that of the no-skill blinded forecast based on haze occurrence frequency. However, the machine still produces a relatively high number of missing forecasts (false negative for haze events), likely owing to its lack of experience in identifying atypical patterns. Nevertheless, this effort has demonstrated a promising prospect of using deep learning algorithms to predict the occurrence of extreme environmental and weather events, and to advance knowledge about these still poorly known phenomena.