Deep Learning Approach for Assessing Air Quality During COVID-19 Lockdown in Quito

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Weather Normalized Models (WNMs) assess urban pollution under business-as-usual scenarios.
• Previous WNM development utilized Machine Learning (ML), but Deep Learning (DL) remained unexplored for quantifying event impacts like COVID-19 lockdowns.

Purpose of the Study:

• To develop novel WNMs using DL algorithms.
• To compare the performance of five DL architectures against a Gradient Boosting Machine (GBM) ML model.
• To quantify the impact of COVID-19 lockdowns on urban air quality in Quito, Ecuador.

Main Methods:

• Proposed WNMs based on five DL architectures: Long-Short Term Memory (LSTM), Bidirectional Recurrent Neural Network (BiRNN), and others.
• Compared DL model performance against a Gradient Boosting Machine (GBM) approach.
• Analyzed concentrations of five air pollutants (CO, NO2, PM2.5, SO2, O3) and identified important predictive features.

Main Results:

• LSTM and BiRNN models demonstrated superior performance as WNMs compared to other DL architectures and GBM.
• Key predictors for air quality included time of day, data index, day of the year, relative humidity, wind speed, and solar radiation.
• COVID-19 full lockdown led to significant pollutant reductions: CO (-48.75%), SO2 (-45.76%), PM2.5 (-42.17%), NO2 (-63.98%).
• NO2 reduction correlated with a +26.54% increase in O3 concentration.

Conclusions:

• DL models, particularly LSTM and BiRNN, are effective for developing advanced WNMs.
• These WNMs accurately quantify the effects of lockdowns on urban air quality.
• Temporal and meteorological factors are crucial for predicting air pollutant concentrations.