Data assimilation


An “intelligent city system” will integrate several simulation models, e.g., for water management, air quality forecasting and analysis, traffic assignment, estimation of city emissions, etc. Unfortunately, these models are often limited by their high uncertainties. At the same time, an increasing amount of observational data is available to the public and decision makers. In particular, low-quality mobile sensors can get deployed in large numbers. There is a clear challenge in combining information from the simulation models and all the available data, either for analysis, real-time evaluation, forecasting or even planning.


For every system of interest, the main goal is estimate its state with the lowest uncertainty. For instance, the chemical state of the atmosphere will include the concentrations of several harmful chemical species in the streets. This state is to be estimated for present time, past times (analyzing) and future times (forecasting). Along with the estimation of the state, uncertainty quantification should be computed. In order to efficiently decrease the uncertainty, one should determine what to observe and at what locations, which is referred to as optimal network design.

State of the art and challenges

There is little work on data assimilation at the urban scale. In air quality, the observations have been employed for model evaluation and source apportionment, but advanced data assimilation and uncertainty quantification have barely been applied in this context. The lack of uncertainty quantification studies is by itself a limiting factor for data assimilation. One challenge is to explore the performance of known methods, if applicable, and to find out their limitations in the context.


The following work is anticipated:

  • Uncertainty quantification of emission inventories and model simulations, with application to air quality and carbon emissions: In order to carry out these tasks, there is a need for simulations with a chemistry-transport model adapted to urban scale. A model like ADMS Urban is well established and may carry out simulations with very fine resolution. A detailed evaluation of the performance of the model and an estimation of its uncertainties may be carried out. The most ambitious study would involve several models (ensemble simulations), and observational data to calibrate the uncertainty quantification. Similar work could be carried out on traffic emission models, which have been scarcely evaluated inside a city like Paris, despite the availability of numerous traffic counters. Such work would also help quantifying the uncertainty of air quality models, since they are highly dependent on emission estimations.
  • Assimilation of air quality observations, so as to improve the simulated maps of NO2, O3 and particulate matter: This part is essentially related to air quality. Using the air quality observations (mainly, NO2, O3, particulate matter) from an observation network, we can dramatically improve the model simulations. The prototype “Votre Air” demonstrated that this can be applied in near real-time. The objective is to better evaluate the exposure of urban population to harmful pollutants.
  • Application to noise pollution, especially with the assimilation of mobile observations: Another key application at urban scale lies in the estimation of population exposure to noise pollution. The assimilation of low-quality data as measured by smartphones microphones (see the SoundCity mobile application) is investigated by CityLab teams.
  • Inverse modeling of carbon emissions at high resolution (down to the street): Using a transport model that simulates the dispersion of carbon emissions, it is possible to quantify the link between the concentrations observed in the air and the emissions. Using techniques for inverse modeling, it may be possible to improve emission inventories. It is a difficult task because of the computational burden of such techniques and of the numerous uncertainties in the simulation models. It is however important to assess how much emission inventories can be improved on the basis of observations of air concentrations.
  • The optimal design of observation networks: The performance of data assimilation strongly depends on the available observations. Observation networks may require huge investments. As a consequence, an important issue is the optimal design of observation networks. This issue is raised when a new network is deployed, but also when, for an existing network, sensors are added, removed or relocated. The underlying question is to maximize the information brought by the network, so that its measurements may optimally improve the models simulations (or the emissions in case of inverse modeling).

Permanent link to this article: