A study on extending the use of air quality monitor data via deep learning techniques

As the environmental awareness of urban citizens increases, traditional air quality monitoring stations cannot satisfy the need for air quality data at high temporal and spatial resolution due to their high construction and maintenance costs. Low-cost air quality monitors are being increasingly used for this purpose because of their portability and affordable price. However, low-cost monitors are usually beset by data quality issues, and the number of mounted air pollutant sensors is limited by the restriction of the cost and size of monitors. Therefore, we propose to extend the use of air quality monitor data via a deep learning technique called long short-term memory (LSTM). The extension is embodied in two aspects: first, calibration of air pollutant concentration data; and second, provision of indicative information about air pollutants where no corresponding sensors are available. A low-cost air quality monitor called KOALA (Knowing Our Ambient Local Air Quality), which was deployed in Sydney (Australia), was used as an example to prove this method's feasibility. Data from a 90-day period were used for model training, and data from a 30-day period were used for model validation. Random forest models were used for selecting the most useful LSTM model input variables. Historical 24-h information was incorporated to improve the performance of the LSTM models. The results showed that: first, LSTM models can be used to calibrate KOALA carbon monoxide (CO) data with the optimum input being raw CO measurements and the corresponding standard deviation information; and second, LSTM models can be used to estimate ozone concentration with the optimum input being CO concentration and three meteorological parameters [i.e. top soil layer temperature, 10 m U wind (earth-relative), and net shortwave radiation flux at the ground] generated through a deterministic model known as WRF (Weather Research and Forecasting). In addition, the LSTM ozone estimation model showed good performance at both the training location and a location 11 km away, indicating that the proposed method can be used to provide indicative information about air pollutants around the training location.