Premium
BLUE‐based NO 2 data assimilation at urban scale
Author(s) -
Tilloy Anne,
Mallet Vivien,
Poulet David,
Pesin Céline,
Brocheton Fabien
Publication year - 2013
Publication title -
journal of geophysical research: atmospheres
Language(s) - English
Resource type - Journals
eISSN - 2169-8996
pISSN - 2169-897X
DOI - 10.1002/jgrd.50233
Subject(s) - covariance , estimator , data assimilation , mean squared error , computation , parameterized complexity , gaussian , scale (ratio) , best linear unbiased prediction , statistics , variance (accounting) , computer science , environmental science , mathematics , algorithm , meteorology , geography , cartography , selection (genetic algorithm) , physics , business , accounting , quantum mechanics , artificial intelligence
We aim at optimally combining air quality computations, from the Gaussian model ADMS Urban, and ground observations at urban scale. An ADMS simulation generated NO 2 concentration fields across Clermont‐Ferrand (France) down to street level, every 3 h for the full year 2008. A monitoring network composed of nine fixed stations provided hourly observations to be assimilated. Every 3 h, we compute the so‐called BLUE (best linear unbiased estimator), which is a concentration field merging ADMS outputs and ground observations. Its error variance is supposed to be minimal under given assumptions regarding the errors on observations and model simulations. A key step lies in the modeling of error covariances between the computed NO 2 concentrations across the city. We introduce a parameterized covariance which heavily relies on the road network. The covariance between two locations depends on the distance of each location to the road network and on the distance between the locations along the road network. Efficient parameters for the covariances are primarily chosen according to prior assumptions, χ 2 diagnosis and leave‐one‐out cross‐validations. According to the cross‐validations, the improvements due to the assimilation seem moderately far from the observation network, but the root mean square error roughly decreases by 30–50% in the main city where the station density is high. The method is computationally tractable for the generation of improved concentration fields over a long period, or for day‐to‐day forecasts.