LAUSR

Abstract Outdoor air pollution remains a major environmental threat to the public, especially those who reside in highly urbanised areas. Recent studies have revealed the effectiveness of early-warning mechanisms that enable the public reduce their exposure to air pollutants. This highlights the need for accurate air quality forecasts. However, air quality in many developing and highly urbanised countries remains unmonitored. Hence, a novel spatiotemporal interpolation modelling approach based on a deep learning and wavelet pre-processing technique was proposed in this paper. In more detail, Long Short-term Memory (LSTM) neural networks and Discrete Wavelet Transformation (DWT) were utilised to model the spatial variability of hourly NO2 levels at six urban sites in Central London, the United Kingdom. The models were trained using only the NO2 concentration data from the neighbouring sites. Benchmark models such as plain LSTM and Multilayer Perceptron (MLP) models were also developed to validate the effectiveness of the proposed models. The proposed wavelet-based spatiotemporal models were found to provide superior forecasting results, explaining 77% to 93% of the variability of the actual NO2 concentration levels at most sites. The overall results reveal the very promising potential of the proposed models for the spatiotemporal characterisation of outdoor air pollution.