The application of artificial neural networks for discrete wavelength retrievals of atmospheric nitrogen dioxide from space

Despite emission reductions in Europe, air quality continues to be a major health and policy issue. Significant areas of uncertainty persist in relating emissions, atmospheric composition and human exposure within the urban atmosphere. Furthermore, air quality continues to worsen in some highly populated parts of the world. The current air quality monitoring framework is based upon bottom-up emission estimates coupled with sparse in situ monitoring. Research at the University of Leicester in the UK is being conducted to investigate the feasibility of using a technique of discrete wavelength sunlight spectroscopy to derive concentrations of the pollutant nitrogen dioxide from a satellite platform. This technique has the potential to enable very light and compact instrumentation and may subsequently provide abundant air quality data of significant value to users and policy makers. A back propagation multi-layered perceptron artificial neural network (ANN) has been developed to retrieve atmospheric slant columns of nitrogen dioxide from simulated measurements. The ANN approach enables retrievals to be performed much faster than other retrieval methods once the network has been appropriately trained, which is a particularly useful feature in instances where a large quantity of retrievals is required in near real time. To generate the required training data for the ANN to understand the necessary relationships a radiative transfer model SCIATRAN was run to provide millions of spectral intensities and slant column concentrations. To enable the radiative transfer simulations to realistically portray urban air quality the SCIATRAN model was fed atmospheric profile and aerosol data from modelled air quality forecasts over London to enable assimilation of the atmospheric composition of a typical urban environment. The training data produced by SCIATRAN was configured to span a range of solar azimuth and zenith angles to provide results which are applicable to all low earth orbit configurations. Once produced the simulated spectra were fed into an instrument model which applied filter functions to discrete wavelength positions within the spectra with various FWHM's and SNR's to cover a range of possible instrument models. The supercomputer at the University of Leicester (ALICE) has been utilised extensively in the network training process, and investigations into the ANN's ability to retrieve nitrogen dioxide from unseen test data of simulated concentrations will be presented.