Air pollution and the emission of GHGs is the main cause of climate change with annual global emission levels still on the rise (Friedlingstein et al., 2019). In particular, anthropogenic GHG emissions from the combustion of fossil fuels in industrial plants or for transportation are harmful to the environment and contribute to global warming trends (Ledley et al., 1999). Besides the primary greenhouse gas, CO, the burning of fossil fuels also emits molecules like NO and CO, which have been used as proxy for the estimation of CO emissions (Berezin et al., 2013). Detailed information about sources and distribution of air pollutants within the atmosphere is of high relevance for a number of applications with climate change impact, including the compilation of emission inventories (Eggleston et al., 2006), the design and implementation of pollution limits (Bollen and Brink, 2014), and the quantification of large anthropogenic emissions (Liu et al., 2020).
At present, continual data on air pollution concentrations in the atmosphere are primarily collected through two different approaches with distinct drawbacks. On the Earth’s surface, networks of measurement stations record the concentration of various chemicals at select locations (Guerreiro et al., 2014). Such networks are commonly run by environmental agencies and provide frequent measurements while often lacking in spatial coverage. This drawback can be partly addressed by space-borne air pollution monitoring: satellites equipped with spectrometers measure the abundance of select molecules in the form of atmospheric column densities (Gupta et al., 2006). While their position in Earth’s orbit allows satellites to frequently map most locations on Earth, remote sensing spectrometers currently only provide spatial resolutions in the kilometer range and with little information about the pollutant’s vertical distribution. Specifically, the estimation of concentrations near the surface, where these pollutants originate from, is a non-trivial task (Scheibenreif et al., 2021). One of the primary anthropogenic air pollutants is Nitrogen Dioxide (NO). Elevated levels of NO harm the vegetation, contribute to acid rain, and act as a precursor of potent GHGs like Ozone (Montzka et al., 2011). Additionally, NO is jointly emitted with CO during the combustion of fossil fuels at high temperatures, making it a suitable proxy to identify CO emission sources (Konovalov et al., 2016; Goldberg et al., 2019). This work leverages a large body of publicly available NO concentration measurements on the ground by the European Environment Agency’s111eea.europa.eu (EEA) network of air quality stations and satellite measurements from the European Space Agency’s (ESA) Copernicus program to investigate the distribution of air pollutants through a deep learning approach. The results of this work enable the identification of major sources of GHG emissions and their temporal monitoring on a global scale.
Various prediction and interpolation techniques have been used to derive detailed information about the spatial distribution of air-borne pollutants such as GHGs. Typically, these models are based on point measurements from air quality monitoring stations that are spatially limited to specific locations. Beyond interpolation with geostatistical approaches like kriging(Janssen et al., 2008), land-use-regression (LUR) is commonly applied to incorporate covariates such as population density or traffic data into the models (see Hoek et al., 2008, for a review). LUR models often involve variable selection procedures to identify predictive inputs over large sets of candidate variables, making it difficult to scale to regions not covered by detailed datasets, even if some air quality measurements are available. Building on existing work that incorporates satellite measurements into LUR frameworks (Novotny et al., 2011), we extend this approach to model air pollution at high spatial resolution solely from satellite data. Our work is based on NO concentration measurements by the EEA. We consider NO as pollutant of interest due to its relevance as major anthropogenic air pollutant and chemical properties that facilitate its detection from space with high accuracy (opposed to GHGs like CO). Additionally, it is co-emitted with CO in the burning of fossil fuels, which makes it possible to constrain CO emissions from NO measurements (Berezin et al., 2013). To facilitate the identification of air pollutant sources, which are commonly located on the ground, we model surface-level concentrations (rather than e.g. atmospheric column densities). The EEA network of air quality stations provides frequent (mostly hourly) measurements of NO concentrations at more than 3,000 locations in Europe. Additionally, remote sensing data from ESA’s Sentinel-2 and Sentinel-5P satellites is utilized to model air quality. Sentinel-2 is a constellation of two satellites carrying the Multi Spectral Instrument, a spectrometer covering the visible, near-infrared and shortwave-infrared wavelengths with imaging resolutions between 10 and 60 meters (Drusch et al., 2012). Sentinel-2 data is widely used in applications like land cover classification or crop monitoring (Helber et al., 2019) but also for the monitoring of GHGs at locations of interest (e.g., based on the presence of smoke plumes, Mommert et al., 2020)
. In our work, globally available and continually updated Sentinel-2 images replace conventional LUR predictor variables such as street networks, population density or vegetation information. The Sentinel-5P satellite observes trace-gases and aerosols in the atmosphere through differential optical absorption spectroscopy(Veefkind et al., 2012). It provides daily global coverage for gases including NO, O, CO or CH with a spatial resolution of km. We utilize the NO tropospheric column density product of Sentinel-5P to model the temporal variation in surface NO levels.
This work approaches the estimation of air pollution as a supervised computer vision problem. We collect a dataset of harmonized remote sensing data from Sentinel-2 and Sentinel-5P, spatially and temporally aligned with measurements from air quality monitoring stations. The proposed model is trained on pairs of remote sensing input and air quality target values (see Fig.1), which yields a system that predicts air pollution levels solely from globally available remote sensing data222code available at github.com/HSG-AIML/RemoteSensingNO2Estimation.
3.1 Data Processing
We consider the 2018-2020 timespan, historically limited by the start of the Sentinel-5P nominal mission. NO measurements by EEA air quality stations are filtered to remove values with insufficient quality (validity or verification value 1). Besides modelling the entire 2018-2020 timespan, we also investigate the possibility to estimate NO concentrations at quarterly and monthly frequencies. To that end, the mean of NO measurements for each frequency is used as prediction target. To build the dataset, we downloaded Sentinel-2 Level-2A data (i.e. corrected for atmospheric effects and enriched with cloud masks) with low cloud-coverage at the locations of air quality stations, containing 12 different bands (band 10 is empty in the case of Level-2A data). The images were then cropped to 120120 pixel size (1.21.2 km) centered at the location of interest, and all bands were upsampled to 10 m resolution with bilinear upsampling. Additionally, we visually inspected the RGB bands of all images to ensure that no clouds or artifacts are present. Similarly, Sentinel-5P data over Europe was downloaded for the 2018-2020 timespan (5449 Level-2 products) and mapped to a common rectangular grid of 0.050.05 (55 km) resolution after removing invalid measurements (qa_value 75). The resulting dataset was averaged at the different temporal frequencies and 2020 km regions at the locations of air quality stations were extracted. To facilitate processing despite the coarse resolution (500 lower than Sentinel-2), we linearly interpolated the Sentinel-5P data to 10 m resolution and cropped to 120120 pixel centered at the locations of interest.
3.2 Model Architecture
). The input layer is modified to accommodate the 12-band Sentinel-2 input data and the final layer is replaced by two dense layers with ReLU activation (namedhead et al., 2019). After pretraining, the final classification layer is replaced by the head, i.e., only the trained convolutional backbone of the ResNet is retained. Intuitively, learned features that are informative for LCC (e.g., distinguishing industrial areas from forests) will also be useful when estimating emission profiles of different areas. To handle additional input data from Sentinel-5P, the model architecture is extended with a small sub-network, consisting of two convolutional layers (with 10,15 channels and kernel sizes 3,53), and a final linear layer. This sub-network is much smaller than the ResNet-50 used to process the Sentinel-2 input stream to reflect the lower native resolution and single band nature of the Sentinel-5P data. It learns a 128 dimensional latent vector from the Sentinel-5P input image. To obtain an NO prediction, the latent vectors of both input-streams are concatenated and again processed by the head
with adjusted input dimensions (2048+128). All presented models were trained 10 times with varying seeds, mean-squared-error loss function and random train/test/validation split of 60:20:20. To limit overfitting, training is stopped once the loss on the validation set stops decreasing. Additionally, we employ random flipping and rotation of the inputs as augmentation during training.
To assess the predictive power of Sentinel-2 images for air pollution prediction we initially train a model on Sentinel-2 images as inputs with air quality station measurements as target. Using only Sentinel-2 images forces the model to associate features that are apparent in medium-resolution satellite imagery, like built-up areas, forests or streets, with representative NO levels. Training this model from scratch leads to a mean-absolute-error (MAE) of 8.060.49 and R2-Score of 0.250.05 (see Table 1), presumably limited by the dataset size of only 3,227 images. Following the intuition that LCC shares predictive features with air pollution prediction, we then investigated a transfer learning approach by pre-training the ResNet backend on BigEarthNet (590,326 images with multi-label annotations, Sumbul et al., 2019). Using the pretrained backend in the NO prediction model and fine-tuning on the Sentinel-2 images at air quality stations, we obtain a significantly better performance. The MAE drops to 6.620.17 with an R2-Score of 0.450.03. This first result supports our hypothesis that medium-resolution satellite imagery is valuable for the estimation of ambient air pollution. We then investigated ways of incorporating tropospheric column density measurements of NO from Sentinel-5P into the model using a second input stream. The additional satellite data results in a further performance increase with an MAE of 5.920.44 and R2-Score of 0.540.04 and allows us to derive detailed pollution maps for any location of interest (see Fig. 2). Inclusion of Sentinel-5P data, which is updated daily, also provides us with a way of modeling temporal variations in NO levels. Aggregating the data at higher frequency significantly increases the number of observations (from 3.1k to 19.6k quarterly and 59.6k monthly samples), which enables the model to maintain a performance comparable to the static predictions (MAE of 6.240.22 and 6.520.15 for quarterly and monthly predictions, respectively). Similarly, the R2-Scores remain at 0.520.05 (quarterly) and 0.510.01 (monthly) despite the increase in prediction frequency. This makes it possible to model seasonal changes in NO concentrations with good accuracy (see Fig. 3).
We present an end-to-end approach for the estimation of surface NO concentrations with deep learning. Utilizing only remote sensing data as inputs, it is possible to model arbitrary regions on Earth, independent of the availability of detailed datasets as commonly used in the prediction of air pollutant distributions. Qualitative evaluation shows that our models are robust across most regions of Europe, except for rare atypical locations that are badly represented in our dataset, e.g., snowy mountain peaks. In future work, measurements from air quality networks outside of Europe can be incorporated into model training to improve model generalization.
The focus of this work on NO allows us to leverage a large corpus of pollutant measurements from air quality stations and from space to better localize the sources of air pollution and GHG emitters. This information enables an approximate analysis of the spatial and temporal distribution of air pollutants and GHG emissions alike, providing constraints that are vital for our effort to reduce GHG emissions and reaching the net-zero emission target.
We thank the ESA Copernicus programme and the European Environment Agency for providing the data used in this work.
- Multiannual changes of CO2 emissions in China: indirect estimates derived from satellite measurements of tropospheric NO2 columns. Atmospheric Chemistry and Physics 13 (18), pp. 9415–9438. Cited by: §1, §2.
- Air pollution policy in Europe: quantifying the interaction with greenhouse gases and climate change policies. Energy Economics 46, pp. 202–215. Cited by: §1.
- Sentinel-2: ESA’s optical high-resolution mission for GMES operational services. Remote Sensing of Environment 120, pp. 25–36. Cited by: §2.
- 2006 IPCC guidelines for national greenhouse gas inventories. Cited by: §1.
- Global carbon budget 2019. Earth System Science Data 11 (4), pp. 1783–1838. Cited by: §1.
- Exploiting OMI NO2 satellite observations to infer fossil-fuel CO2 emissions from US megacities. Science of The Total Environment 695, pp. 133805. Cited by: §1.
- Air quality status and trends in Europe. Atmospheric Environment 98, pp. 376–384. Cited by: §1.
- Satellite remote sensing of particulate matter and air quality assessment over global cities. Atmospheric Environment 40 (30), pp. 5880–5892. Cited by: §1.
Deep residual learning for image recognition.
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778. Cited by: §3.2.
- Eurosat: A novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12 (7), pp. 2217–2226. Cited by: §2.
- A review of land-use regression models to assess spatial variation of outdoor air pollution. Atmospheric Environment 42 (33), pp. 7561–7578. Cited by: §2.
- Spatial interpolation of air pollution measurements using CORINE land cover data. Atmospheric Environment 42 (20), pp. 4884–4903. Cited by: §2.
- Estimation of fossil-fuel CO2 emissions using satellite measurements of ”proxy” species. Atmospheric Chemistry and Physics 16 (21), pp. 13509–13540. Cited by: §1.
- Climate change and greenhouse gases. Eos, Transactions American Geophysical Union 80 (39), pp. 453–458. Cited by: §1.
- A methodology to constrain carbon dioxide emissions from coal-fired power plants using satellite observations of co-emitted nitrogen dioxide. Atmospheric Chemistry and Physics 20 (1), pp. 99–116. Cited by: §1.
- Characterization of industrial smoke plumes from remote sensing data. Cited by: §2.
- Non-CO2 greenhouse gases and climate change. Nature 476 (7358), pp. 43–50. Cited by: §1.
- National satellite-based land-use regression: NO2 in the United States. Environmental Science & Technology 45 (10), pp. 4407–4414. Cited by: §2.
- A Novel Dataset and Benchmark for Surface NO2 Prediction from Remote Sensing Data Including COVID Lockdown Measures. In IGARSS 2021-2021 IEEE International Geoscience and Remote Sensing Symposium, Cited by: §1.
- Bigearthnet: a large-scale benchmark archive for remote sensing image understanding. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium, pp. 5901–5904. Cited by: §3.2, §4.
- TROPOMI on the ESA Sentinel-5 Precursor: a GMES mission for global observations of the atmospheric composition for climate, air quality and ozone layer applications. Remote Sensing of Environment 120, pp. 70–83. Cited by: §2.
Appendix A Appendix