A temporal-spatial interpolation and extrapolation method based on geographic Long Short-Term Memory neural network for PM2.5

Nowadays, real-time air pollution monitoring has been an important approach for supporting pollution control and reduction. However, due to the high construction cost and limited detection range of monitoring stations, not all the air pollutant concentrations in every corner can be monitored, and a whole picture of the spatial distribution of air pollution is usually lacked for comprehensive spatial analysis and air quality control. To address this problem, satellite remote sensing and spatial interpolation/extrapolation technologies have been commonly used in past research. However, the spatial distribution calculated by remote sensing techniques could be less accurate due to the limited amount of recorded data for testing and adjustments. Performance of traditional spatial interpolation/extrapolation techniques, such as Kriging and IDW, was limited by several subjective assumptions and pre-set formulations that are less suitable for non-linear real-world situations. As an alternative, machine learning and neural network-based methods have been proposed recently. However, most of these methods failed to well consider the long short temporal trend and spatial associations of air pollution simultaneously. To overcome these limitations, this paper proposed a newly designed spatial interpolation/extrapolation methodology namely Geo-LSTM to generate the spatial distribution of air pollutant concentrations. The model was developed based on the Long Short-Term Memory (LSTM) neural network to capture the long-term dependencies of air quality. A geo-layer was designed to integrate the spatial-temporal correlation from other monitoring stations. To evaluate the effectiveness of the proposed methodology, a case study in Washington state was conducted. The experimental results show that Geo-LSTM has a RMSE of 0.0437, and is almost 60.13% better than traditional methods like IDW.