ResNet-LSTM for Real-Time PM<sub>2.5</sub> and PM₁₀ Estimation Using Sequential Smartphone Images
Shiguang Song, Jacqueline C. K. Lam, Yang Han, Victor O. K. Li
Abstract
Attempts have been made to estimate PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> and PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sub> values from smartphone images, given that deploying highly accurate air pollution monitors throughout a city is a highly expensive undertaking. Departing from previous machine learning studies which primarily focus on pollutant estimation based on single day-time images, our proposed deep learning model integrates Residual Network (ResNet) with Long Short-Term Memory (LSTM), extracting spatial-temporal features of sequential images taken from smartphones instead for estimating PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> and PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sub> values of a particular location at a particular time. Our methodology is as follows: First, we calibrated two small portable air quality sensors using the reference instruments placed in the official air quality monitoring station, located at Central, Hong Kong (HK). Second, we verified experimentally that any PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> and PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sub> values obtained via our calibrated sensors remain constant within a radius of 500 meters. Third, 3024 outdoor day-time and night-time images of the same building were taken and labelled with corresponding PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> and PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sub> ground truth values obtained via the calibrated sensors. Fourth, the proposed ResNet-LSTM was constructed and extended by incorporating meteorological information and one short path. Results have shown that, as compared to the best baselines, ResNet-LSTM has achieved 6.56% and 6.74% reduction in MAE and SMAPE for PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> estimation, and 13.25% and 11.03% reduction in MAE and SMAPE for PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sub> estimation, respectively. Further, after incorporating domain-specific meteorological features and one short path, Met-ResNet-LSTM-SP has achieved the best performance, with 24.25% and 20.17% reduction in MAE and SMAPE for PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.</sub> 5 estimation, and 28.06% and 24.57% reduction in MAE and SMAPE for PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sub> estimation, respectively. In future, our deep-learning image-based air pollution estimation study will incorporate sequential images obtained from 24-hr operating traffic surveillance cameras distributed across all parts of the city in HK, to provide full-day and more fine-grained image-based air pollution estimation for the city.