Synergistic data fusion of satellite observations and in-situ measurements for hourly PM2.5 estimation based on hierarchical geospatial long short-term memory

PM2.5 as a primary air pollutant has adverse effects on the environment and public health. The air quality monitoring stations are distributed sparsely and unevenly, making it difficult to provide continuous and precise regional measurements, which can be supplemented by satellite observations. However, most satellite-based approaches for air pollution estimation are difficult to extract the spatio-temporal dependencies effectively, leading to lower accuracy in long-term prediction and assessment of episodic changes. To fill this gap, a hierarchical geospatial long short-term memory method (HG-LSTM) by considering the geospatial autocorrelation was proposed for hourly PM2.5 estimation with 2-km spatial resolution in Yangtze River Delta (YRD) urban agglomeration. The superior accuracy of the HG-LSTM is compared with other models via the site-based and year-based cross-validation (CV) tests, indicating geospatial autocorrelation exerts non-negligible impacts on the PM2.5 estimation. The estimations are consistent with the in-situ observations with site-based CV R2 of 0.88. The deviations less than 10 μ g/m3 account for over 80%. The PM2.5 spatiotemporal characteristics in the YRD reveal that PM2.5 concentrations are higher in the morning and decline significantly in the afternoon. As well, elevated PM2.5 values are accumulated in the northern regions of the study area. Although the prediction accuracy decreases as the augment of prediction timesteps, the results can still be useful to detect air pollution changes in the near future. Overall, the HG-LSTM model can estimate hourly PM2.5 concentrations accurately and seamlessly, which is beneficial for air pollution monitoring and environmental protection strategy formation.