Uncertainty-Aware Deep Learning Architectures for Highly Dynamic Air Quality Prediction

Abstract

Forecasting air pollution is considered as an essential key for early warning and control management of air pollution, especially in emergency situations, where big amounts of pollutants are quickly released in the air, causing considerable damages. Predicting pollution in such situations is particularly challenging due to the strong dynamic of the phenomenon and the various spatio-temporal factors affecting air pollution dispersion. In addition, providing uncertainty estimates of prediction makes the forecasting model more trustworthy, which helps decision-makers to take appropriate actions with more confidence regarding the pollution crisis. In this study, we propose a multi-point deep learning model based on convolutional long short term memory (ConvLSTM) for highly dynamic air quality forecasting. ConvLSTM architectures combines long short term memory (LSTM) and convolutional neural network (CNN), which allows to mine both temporal and spatial data features. In addition, uncertainty quantification methods were implemented on top of our model's architecture and their performances were further excavated. We conduct extensive experimental evaluations using a real and highly dynamic air pollution data set called Fusion Field Trial 2007 (FFT07). The results demonstrate the superiority of our proposed deep learning model in comparison to state-of-the-art methods including machine and deep learning techniques. Finally, we discuss the results of the uncertainty techniques and we derive insights.

A generic framework for monitoring pollution plumes in emergencies using UAVs

“…The performance of our deep learning model has been evaluated and compared to several state-of-the-art methods through extensive experiments in an anterior work [6] and the effectiveness of the proposed model has been demonstrated.…”

“…This allows to get intervals that bracket the real values. To that end, we have compared the quantile regression method with Monte Carlo dropout technique on top of our forecasting architecture to generate PIs in [6]. In this work we consider uncertainty based on quantiles since the latter approach is not computationally expensive and generates good quality PIs.…”

“…Therefore, as a high coverage level (e.g. 80%) is sufficient to encompass the real observations within the PIs most of the time [6], the width could reflect the uncertainty.…”

“…It's worth to mention that several coverage levels for the PIs were tested ranging from 10% to 90%, and a coverage level of 80% is chosen here as it generates PIs with a good compromise between percentage coverage and interval width [6].…”

Deep learning techniques applied to super-resolution chemistry transport modeling for operational uses

“…Machine learning techniques have also emerged as relevant solutions to forecast air quality at stations and using observations (meteorology, concentrations, emissions, landcover, etc.) as predictor variables [7][8][9][10][11][12][13][14][15]. Recently, deep learning schemes based on neural networks (NN) [16] has become more and more popular with increasing computer power and training data availability.…”