Fig. 3: Four representative events of air pollution presented by the predicted surface grids.
From: Improving air quality assessment using physics-inspired deep graph learning
a–c The predicted PM10 surfaces on a sandstorm event in North China (c) of April 15, 2015 and the day (a) before it, and the difference (b) between the two. d–f The predicted PM2.5 surfaces on a haze event in East China (f) of December 23, 2016 and the day (d) before it, and the difference (e) between the two. g–i The predicted O3A8 surfaces on a severe ozone event in Beijing (i) of July 1, 2017 and the day (g) before it, and the difference (h) between the two. j–l The predicted NO2 surfaces on a haze event in Shanghai (l) of December 19, 2018 and the day (j) before it, and the difference (k) between the two. The blue arrow is mean ground wind vector (m/s) and the black isolines present GPH in meter at 500 hPa. The dash lines in (b, e, h, k) show the HYSPLIT clusters of backward air trajectories at 500 m with the percentage.
