Julien Chimot

In the last decades, significant progress was achieved in global monitoring of the atmospheric composition from space, leading to key discoveries on the ozone layer, air quality and climate change. Recent research also shows that satellite observations can be used to monitor the effectiveness of policy measures on air quality and climate change. Therefore, it becomes more and more important to provide accurate data under any atmospheric condition. Furthermore, developing algorithms that can meet the requirements of an operational 24/7 data stream becomes challenging due to the tremendous growth of satellite measurements. Europe in general, the Netherlands specifically, have heavily invested in satellite instruments monitoring short-lived and greenhouse gases. Most notably, one can cite the Dutch-Finnish Ozone Monitoring Instrument (OMI) on-board the NASA’s Earth Observation System (EOS) Aura platform, and the recently launched TROPOspheric Monitoring Instrument (TROPOMI), developed by ESA and the Netherlands, on the Copernicus Sentinel-5 Precursor.

Nitrogen dioxide NO2 and particulate matter (referred to as aerosols) are both important constituents of the atmospheric composition. In the troposphere, NO2 is emitted by fossil fuel combustion, biomass burning, as well as by lightning. It has not only adverse health effects, but also affects our atmosphere: 1) NO2 plays a key role in the formation of tropospheric ozone O3, a toxic element for humans and plants, 2), and NO2 chemistry reactions lead to secondary aerosol formation.

Atmospheric aerosols are particles suspended in the air. Their sources are very mixed. Aerosols can be man-made or natural: e.g. smoke, desert dust, sea spray, nitrates and sulfates. Some of them are directly released in the air as particles: e.g. desert dust. Others are formed in the atmosphere resulting from chemistry reactions by precursor gases, such as SO2 and NO2. Because aerosols scatter and absorb sunlight, they perturb the radiative energy balance of the Earth, and thus affect climate. They also play a key role in the formation of clouds and precipitation. Aerosol effects are recognized as one of the largest uncertainties in our knowledge on climate change. Additionally, the small particles can deeply penetrate into the human respiration system leading to strong adverse health effects.

Both NO2 and aerosols (e.g. nitrates, sulfates) are formed from combustion processes. Because their lifetime in the troposphere is comparable (NO2 typically less than one day, aerosol 1 day to a week), the column concentrations show spatial correlation over regions where the aerosol type is dominated by large urban and industrial activities. The aerosol effects on the sunlight modify the shortwave radiation field in the atmosphere. As a consequence, they are a significant error source when exploiting satellite measurements devoted to trace gases, ocean color and vegetation. Indeed, such measurements are based on the backscattered sunlight at the top of the atmosphere in the shortwave spectrum. This mainly holds for cloud-free conditions where the aerosol signal is dominant.

The main objective of this thesis is to design a new aerosol layer height retrieval in order to improve the operational NO2 retrieval, both in the troposphere, from space-borne instruments for highly polluted events and under cloud-free conditions. This thesis focuses on the exploitation of the OMI satellite measurements acquired in the visible wavelength range (405-490 nm). In addition, we develop numerical methods and tools (e.g. machine learning) in order to support the operational processing of big data amounts from the forthcoming new-generation satellite instruments for air quality and climate research.

In Chap. 2 of this thesis the error in the retrieved OMI tropospheric NO2 vertical column density (VCD) is quantified for current retrieval algorithms and over cloud-free scenes. Ignoring the aerosol effects in cloud-free satellite measurements, i.e. assuming aerosol-free conditions, this leads to a bias in the range of −60% to 20%, in case of high concentrations, Aerosol Optical Thickness (AOT) (550 nm) > 0.6, scattering particles, and for summertime conditions. This clearly shows that a correction for aerosol effects is necessary. In the reference OMI tropospheric NO2 dataset, named DOMINO-v2, an aerosol correction is included through the effective cloud parameters (fraction and pressure) derived by the OMI OMCLDO2 cloud algorithm. These parameters assume a Lambertian reflector model, i.e. clouds are represented as opaque layers that partly cover the observed scene. In the absence of clouds and in the presence of high aerosol loadings, the OMI cloud retrieval algorithm is sensitive to aerosol properties. This leads to an implicit aerosol correction. However, we found that in the original OMCLDO2 algorithm, as used in DOMINO v2, too coarse sampling of the look-up-tables (LUTs) leads to a too low effective cloud pressure retrieval. Consequently, retrieved tropospheric NO2 VCDs are underestimated in the range of −40% to −20%, when aerosol particles are located at high altitude (> 1.5 km), AOT(550 nm) > 0.6, and in summertime conditions (chapter 2). Following this issue, the updated OMCLDO2 algorithm solves this issue.

The aerosol effect on the tropospheric NO2 retrieval depends not only on the vertically integrated aerosol properties, such as AOT, but also how aerosols and NO2 are vertically distributed in the troposphere. When aerosols are located above the NO2 bulk, this leads to a reduction of the OMI measurement sensitivity to NO2. On the contrary, an aerosol layer located below the NO2 bulk leads to an enhancement.

In Chap. 3 of this thesis, we present a novel retrieval technique to derive the Aerosol Layer Height (ALH) from the OMI 477 nm (visible) O2-O2 spectral band over cloud-free scenes, as well over land as water surfaces. We make use of neural network, a specific machine learning approach that has the advantage of being able to process the complete OMI dataset with a low time consumption. Therefore, this algorithm can deal with the big-data challenges of next generation satellite instruments. This algorithm can be used autonomously with the OMI data alone, or also benefit from the synergy with the NASA MODerate resolution Imaging Spectroradiometer (MODIS) instrument, on-board the EOS Aqua. Both OMI-Aura and MODIS-Aqua fly together in the NASA A-Train constellation. The advantage of such a synergy on the ALH retrieval accuracy is twofold: 1) filtering cloud-free OMI observation, and 2) using the reference AOT from the MODIS aerosol products as a prior information for the OMI ALH retrieval.

The performance of the OMI ALH algorithm is assessed by comparing with the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) aerosol observations: first with its related climatology dataset over North-East Asia (Chap 3), and secondly on diverse selected cases (chapter 4). All the comparisons show differences between OMI and CALIOP ALH to be smaller than 800 m in case of cloud-free continental polluted cases and AOT(550 nm) > 0.5. In addition, OMI ALH demonstrates the capability of OMI visible measurements to probe the entire smoke layers, including large loading of absorbing particles, produced by intense biomass fires such as in South-America and East Russia (Chap. 4). The machine learning approach leads to the benefits of a low computing time (apart from the training task). Consequently, large ALH datasets are produced in this thesis, including 3 years of retrievals over North-East China (2005-2007) and a global yearly retrieval for 2006. The latter is obtained using a single processor during approximately 12 hours (Chap. 5). The OMI ALH neural network algorithm shows the potential to exploit the O2-O2 visible spectral band instead of (or in addition to) the traditional and more used O2-A near infrared band from satellite sensors. The main limitations are due to the weak absorption of the O2-O2 complex dimer, the forward aerosol model uncertainty, potential cloud residuals, and the accuracy of the employed surface albedo database. The advantages of the visible O2-O2 band over the near-infrared O2-A band are the higher AOT in the visible and the lower surface albedo over land surfaces.

In Chap. 6 of this thesis, we evaluate the improvements in the tropospheric NO2 VCD retrieval thanks to different aerosol correction schemes using the new developed algorithms based on the OMI O2-O2 visible band: 1) the implicit aerosol correction using effective cloud parameters from the updated OMCLDO2, and 2) an explicit aerosol correction based on the OMI ALH retrieval and other selected aerosol parameters. The evaluation is done by comparing with the old OMCLDO2 algorithm. For that purpose, we reprocess 2 years of cloud-free DOMINO-v2 NO2 data (2006-2007), in summer and winter, over north-East China and South-America. The new implicit aerosol correction shows an improved accuracy in the range of 0-20% for scenes with high aerosol loadings, scattering particles, and summertime conditions. However, such an approach still remains limited in case of more absorbing particles and does not comprehensively represent the single and multiple scattering effects inherent to aerosols. Applying an explicit correction using the OMI ALH retrieval also leads to an improved accuracy on the tropospheric NO2 VCD. It applies a more physical modelling of the aerosol effects on the average light path thanks to the assumed aerosol model in the training dataset of the NN algorithm. Therefore, more realistic vertical averaging kernel are derived, provided that correct aerosol parameters are assumed in the forward model. Higher VCD values are generally obtained by using the OMI ALH compared to OMCLDO2, between 20% and 40% depending on the seasons, regions and pollution episodes. This is likely due to differences between the considered models (i.e. Lambertian opaque reflector for the effective clouds vs. Henyey-Greenstein scattering phase function for aerosols), and/or the assumed height of aerosols and the effective cloud. Finally, the explicit aerosol correction shows higher accuracy in presence of absorbing particles, such as smoke. Overall, its quality does not only depend on the ALH accuracy, but also the set of assumed aerosol parameters (e.g. AOT, single scattering albedo, vertical profile shape, size), surface reflectance, and their resulting combination.

To develop a high-quality performance explicit aerosol correction by using the OMI ALH retrieval, several challenges need to be addressed. The main recommendations are (Chap. 7): to improve the accuracy of the OMI ALH, in particular with respect to the aerosol type knowledge, to define a consistent and accurate set of aerosol parameters that can properly be combined with the retrieved ALH, to pay attention to the OMI radiance closure budget, to improve the accuracy of surface albedo or reflectance and the NO2 vertical profile shape.

Finally, although we mainly use OMI data for this research, all these developments and results can, in principle, be extended to other current and future satellites instruments, like TROPOMI, Sentinel-4 and Sentinel-5 sounders, and also to other trace gas retrievals, such as tropospheric SO2 and HCHO. However, they will have to be adapted to the specificities of this new generation of instruments: e.g. the improvement in the spatial resolution. The small pixel sizes will clearly bring additional challenges such as the 3D effects of clouds, and thus the impacts on the adjacent observation pixels.