Multi-scale three-dimensional variational data assimilation for high-resolution aerosol observations: Methodology and application

Abstract

With an increasing number of air quality monitoring stations installed around the Chinese mainland, high-resolution aerosol observations become available, allowing improvements in air pollution monitoring and aerosol forecasting. However, the multi scales (especially small-scale) information included in high-resolution aerosol observations could not be effectively utilized by the traditional three-dimensional variational method (3DVAR). This study attempted to extend the traditional 3DVAR to a multi-scale 3DVAR with two iteration steps, two-scale-3DVAR (TS-3DVAR), to improve the effectiveness of assimilating high-resolution observations. In TS-3DVAR, the large-scale and small-scale components of observation information were decomposed from the original high-resolution observations using a Gaussian smoothing method and then assimilated using the corresponding large-scale or small-scale background error covariances which were derived from the partitioned background error samples. The data assimilation (DA) analysis field generated by TS-3DVAR is more accurate than 3DVAR in reproducing the field’s multi-scale characteristics, which could thus be used as the initial chemical field of the air quality model to improve aerosol forecasting. Particulate matter with an aerodynamic diameter of less than 2.5 μm (PM2.5) and 10.0 μm (PM10) from the surface air quality monitoring stations from November 01 to November 30, 2018 at 00:00 were assimilated daily to verify the effects of TS-3DVAR and 3DVAR on the aerosol analysis and forecast accuracy. The results showed that TS-3DVAR better constrained both large-scale and small-scale, especially the spatial wavelengths in a range of 54–216 km and those above 351 km. The average power spectra of the TS-3DVAR assimilation increment in the two wavelength ranges were 71.70% and 35.33% higher than those of 3DVAR. As a result, the TS-3DVAR was more effective than 3DVAR in improving the accuracy of the initial chemical field, and thereby the forecasting capability for PM2.5. In the initial chemical field, the 30-day average correlation coefficient (Corr) of PM2.5 of TS-3DVAR was 0.052 (6.12%) higher than that of 3DVAR, and the root mean square error (RMSE) of TS-3DVAR was 3.446 μg m−3 (16.4%) lower than that of 3DVAR. For the forecasting capability for PM2.5 mass concentration, the 30-day average Corr of TS-3DVAR during the 0–24 hour forecast period was 0.025 (5.08%) higher than that of 3DVAR, and the average RMSE was 2.027 μg m−3 (4.85%) lower. The positive effect of TS-3DVAR on the improvement of forecasting capability can last for more than 24 h.