Is bigger always better? Deep learning applications in air quality research
Tuesday, June 29, 2021 1:35 PM to 1:55 PM · 20 min. (Africa/Abidjan)
Stream#2
Information
Contributors:
Abstract:
Forecasting air quality has been one of the early environmental applications of machine learning. In the European IntelliAQ project we explore machine learning to systematically address the challenges of multi-scale spatiotemporal correlations, asymmetric data distributions, and sparse coverage of ground-based station observations. These aspects are also relevant for machine learning from weather and climate data. In recent years very large neural networks have become computationally tractable and several deep learning techniques have been applied to different problems in the context of air quality data analysis and forecasting. The applications range from automated processing of air quality sensor data to the support of numerical model simulations, for example in data assimilation and post-processing. We have tested different interpolation and forecasting approaches from small, out-of-the-box machine learning methods to sophisticated, computationally expensive deep learning networks used for video prediction and we discuss their strengths and limitations.
Abstract:
Forecasting air quality has been one of the early environmental applications of machine learning. In the European IntelliAQ project we explore machine learning to systematically address the challenges of multi-scale spatiotemporal correlations, asymmetric data distributions, and sparse coverage of ground-based station observations. These aspects are also relevant for machine learning from weather and climate data. In recent years very large neural networks have become computationally tractable and several deep learning techniques have been applied to different problems in the context of air quality data analysis and forecasting. The applications range from automated processing of air quality sensor data to the support of numerical model simulations, for example in data assimilation and post-processing. We have tested different interpolation and forecasting approaches from small, out-of-the-box machine learning methods to sophisticated, computationally expensive deep learning networks used for video prediction and we discuss their strengths and limitations.
