Deep Multi-Task Learning Based Urban Air Quality Index Modelling

Abstract

Obtaining comprehensive air quality information can help protect human health from air pollution. Existing spatially fine-grained estimation methods and forecasting methods have the following problems: 1) Only a part of data related to air quality is considered. 2) Features are defined and extracted artificially. 3) Due to the lack of training samples, they usually cannot achieve good generalization performance. Therefore, we propose a deep multi-task learning (MTL) based urban air quality index (AQI) modelling method (PANDA). On one hand, a variety of air quality-related urban big data (meteorology, traffic, factory air pollutant emission, point of interest (POI) distribution, road network distribution, etc.) are considered. Deep neural networks are used to learn the representations of these relevant spatial and sequential data, as well as to build the correlation between AQI and these representations. On the other hand, PANDA solves spatially fine-grained AQI level estimation task and AQI forecasting task jointly, which can leverage the commonalities and differences between these two tasks to improve generalization performance. We evaluate PANDA on the dataset of Hangzhou city. The experimental results show that our method can yield a better performance compared to the state-of-the-art methods.