The synergy between deep learning and numerical modeling in estimating NOx emissions at a fine spatiotemporal resolution

doi:10.21203/rs.3.rs-3129355/v1

Download PDF

Article

The synergy between deep learning and numerical modeling in estimating NO_x emissions at a fine spatiotemporal resolution

https://doi.org/10.21203/rs.3.rs-3129355/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

This study focused on the remarkable applicability of deep learning (DL) together with numerical modeling in estimating NO_x emissions at a fine spatiotemporal resolution during the summer of 2017 over the contiguous United States (CONUS). We employed the partial convolutional neural network (PCNN) and the deep neural network (DNN) to fill gaps in the OMI tropospheric NO₂ column and estimate the daily proxy surface NO₂ map at a spatial resolution of 10 km × 10 km, showing high capability with strong correspondence (R: 0.92, IOA: 0.96, MAE: 1.43). Subsequently, we conducted an inversion of NO_x emissions using the Community Multiscale Air Quality (CMAQ) model at 12 km grid spacing to gain a comprehensive understanding of the chemical evolution. Compared to the prior emissions, the inversion indicated higher NO_x emissions over CONUS (3.21 ± 3.34 times), effectively mitigating the underestimation of surface NO₂ concentrations with the prior emissions. Incorporating the DL-estimated daily proxy surface NO₂ map yielded primary benefits, reducing bias (-1.53 ppb to 0.26 ppb) and enhancing day-to-day variability with higher correspondence (0.84 to 0.92) and lower error (0.48 ppb to 0.10 ppb) across CONUS.

Earth and environmental sciences/Environmental sciences

Earth and environmental sciences/Planetary science

Nitrogen oxides (NO_x), including nitric oxide (NO) and nitrogen oxide (NO₂), are one of the key air pollutants in the troposphere that play a direct role in the formation of harmful components like ozone (O₃) and particulate nitrate (Lei and Wang, 2014; Martin, 2003). Inevitably biased NO_x emissions, resulting from the imperfect classification of emission sources and incorrect emission factors, cause difficulties in monitoring and understanding the impact of NO_x and its products in the atmosphere. For that reason, inversion approaches have been previously employed to constrain NO_x emissions using in-situ and satellite column measurements (Chai et al., 2009; Chu et al., 2005; Goldberg et al., 2019; Jung et al., 2022a, 2022b; Konovalov et al., 2006; Kurokawa et al., 2009; Lamsal et al., 2011; Souri et al., 2016).

Despite significant technological advancements in measurements (e.g., detection, retrieval algorithm, and noise reductions) carried out in the last decades, fundamental characteristics of the measurement data impose limitations on estimating surface NO_x emissions at a fine spatiotemporal resolution. The short lifetime of NO₂ leads to spatial inhomogeneity, making it challenging to infer from in-situ measurements with limited spatial coverage, which are more suitable for small urban areas or point sources. Remote sensing images offer broader spatial coverage; however, they often suffer from missing values due to cloud cover and surface reflectivity (Lops et al., 2021), resulting in insufficient information to estimate emission at a fine temporal resolution. Additionally, in conjunction with a significant reduction in surface NO_x emissions, NO₂ column measurements do not always reflect near-surface air quality when background sources of NO_x, primarily from lightning and soil, make a substantial contribution to the NO₂ column (Jiang et al., 2018; Jung et al., 2022a; Silvern et al., 2019).

In recent years, there have been significant advancements in machine learning and deep learning (DL) algorithms. Efforts have also been directed toward imputing missing data and estimating surface concentrations or column densities of air pollutants such as NO₂ (Ghahremanloo et al., 2021b; Li and Wu, 2021; Yeganeh et al., 2018), PM_2.5/AOD (Ghahremanloo et al., 2022, 2021a; Liu et al., 2019b, 2019a; Lops et al., 2021; Park et al., 2020; Wei et al., 2020, 2019; Xu et al., 2015), and ozone (Di et al., 2017; Ghahremanloo et al., 2023a; Liu et al., 2020; Nabavi et al., 2021; Ren et al., 2020; Requia et al., 2020). Recent studies have demonstrated the effectiveness of deep neural network (DNN) with multiple hidden layers designed to extract high-level features from input data for estimating and predicting air quality (Ghahremanloo et al., 2023). The robust performance of DL-estimated pollutant concentrations enables their application in adverse health impact assessment and emission adjustment, which are highly dependent on the availability and quality of observations.

In this study, our aim is to highlight the significant applicability of DL in estimating NO_x emissions and its synergy with numerical modeling to overcome deficiencies in observation data and inherent limitations in estimating NO_x emissions at a fine spatiotemporal resolution. We utilized the partial convolutional neural network (PCNN) approach to fill the gaps in the Ozone Monitoring Instrument (OMI) tropospheric NO₂ column density (TCDNO₂). Subsequently, we integrated the imputed TCDNO₂ and other predictor variables into a DNN model to estimate the comprehensive map of the surface NO₂ level. This enabled the inversion of NO_x emission at a fine spatiotemporal resolution during the summer of 2017 across the contiguous United States (CONUS).

Estimation of Daily Complete Proxy Surface NO₂ Map and its Validation

The performance of the PCNN model in imputing the OMI TCDNO₂ was validated, as depicted in Fig. 1. We set the averaged OMI TCDNO₂ over the study period (with zero missing value) as the true value. We then applied daily missing data masks on the true value for the spatiotemporal imputation by the PCNN model to reconstruct the averaged OMI TCDNO₂ at each instance. In Figure S1, we presented the spatial distributions of the OMI TCDNO₂ and its imputed results for specific dates (June 1, July 11, and August 8 of the year 2017). The second row of Figure S1 indicated that the PCNN model not only predicted missing values of the OMI TCDNO₂ but also made slight modifications to the original values, resulting in some degree of blurring (Lops et al., 2021). Consequently, we evaluated the distortion of the original image by the PCNN in Fig. 1, which still demonstrated a remarkable agreement with the true value (0.98 of the index of agreement (IOA)). Furthermore, the imputation of missing data exhibited a strong correlation (0.94 of correlation coefficient (R)) with low mismatch (1.93 × 10¹⁴ molecules/cm² of root mean square error (RMSE)), indicating its significant performance in imputing OMI TCDNO₂ and its beneficial impact on estimating a complete proxy surface NO₂ map using the DNN model in conjunction with other predictor variables.

The results of Variance Inflation Factor (VIF) and Shapley Additive Explanations (SHAP) feature performance can be found in Table S1 and Figure S2, respectively. The VIF indexes (Table S1) demonstrated values below 10, indicating that the predictor variables used to train the DNN model did not exhibit significant multicollinearity, which helps prevent model overfitting. In Figure S2, the importance of each predictor variable in estimating surface NO₂ is ranked. The top three variables, in addition to the imputed OMI TCDNO₂, were road density, population density, and percentage of urban space. These variables are known to provide insight into combustion-related emissions in urban areas (Ghahremanloo et al., 2023b, 2022; Schneider and van der A, 2012).

Figure 2 presents the spatial distribution of the DL-estimated complete proxy surface NO₂ map for the summer of 2017 (June to August). It also includes the results of 10-fold cross-validation using NO₂ concentrations from the Environmental Protection Agency (EPA) Air Quality System (AQS) at 18 UTC, over the CONUS, urban and rural regions. The 10-fold cross-validation method is commonly employed to assess the performance of machine learning models. It involves dividing the data into ten random groups and conducting multiple tests, with nine groups used for training and one group for validation. The map in Fig. 2 (a) illustrates elevated NO₂ concentrations dispersed over metropolitan areas such as Seattle, Los Angeles, San Francisco, Salt Lake City, Chicago, and New York (4.30 ± 4.17 ppb), attributed to significant primary sources from transportation and industrial activity. In comparison, Rural areas exhibit relatively lower values (1.05 ± 1.04 ppb). The results of the 10-fold cross-validation demonstrate the strong performance of the DNN model in estimating surface NO₂ concentrations and high spatial inhomogeneity, with R values ranging from 0.85 to 0.92, IOA values ranging from 0.91 to 0.96, and mean absolute error (MAE) values from 0.81 to 1.66 ppb. The accuracy of estimated proxy surface NO₂ concentrations in rural regions is slightly lower than that of urban areas, likely due to the limited number of AQS sites in rural areas, as the AQS primarily targets pollutants in urban environments. The data deficiency in rural regions can also be linked to the limited number of predictor variables used in training the DNN model, highlighting the significance of the spatiotemporal imputation of OMI TCDNO₂ using the PCNN model. This imputation provides ample non-missing data over the area where operational monitoring sites are spatially scattered and sparse.

Inversion of NO_x Emissions Using Daily Complete Proxy Surface NO₂ Map

In order to demonstrate the benefits of using the DL-estimated complete proxy surface NO₂ map in not only alleviating the uncertainty of NO_x emissions but also promoting a comprehensive understanding of the transport and chemical evolution of NO_x, we adjusted NO_x emissions with the WRF-CMAQ modeling system by applying iterative finite difference mass balance (IFDMB) approach. This approach calculates an optimal mass balance solution by accounting for the non-linear relationship between NO₂ column density and NO_x emissions (Cooper et al., 2017; Lamsal et al., 2011; Mun et al., 2023).

Figure 3 (a) to (c) compares the spatial distributions of averaged surface NO₂ concentrations estimated by the DNN model and simulated by CMAQ with the prior and posterior NO_x emissions in the summer of 2017. Surface NO₂ concentrations with the prior NO_x emissions were significantly underestimated compared to the complete proxy surface NO₂ map (Fig. 3 (b)). The underestimation was observed in both urban (42.70%) and rural areas (66.53%), which was consistent with the comparison to satellite column measurement (Jung et al., 2022a). This significant underestimation in the surface NO₂ concentrations was greatly mitigated when the posterior emissions were used. The spatial distributions of the prior and posterior NO_x emissions are shown in Figure S3. Additionally, the prior and posterior NO_x emissions and their ratio for the states of the US are given in Table S2. In general, the ratio between the prior and posterior NO_x emissions was 3.21 ± 3.34 over the CONUS. Moreover, we found that inversion suggests relatively larger increases in the less populated states (< 0.3 Kg/day of the prior NO_x emissions) by 46.36%.

To quantify the improvement in the modeled surface NO₂ concentrations as a benefit of using surface daily complete proxy surface NO₂ map estimated by DL, we displayed the change in the daytime averaged bias of surface NO₂ concentrations between the model output with prior and posterior emissions and the AQS measurement in Fig. 3 (d). The cool color indicates a reduction in bias, while warm colors indicate an increase in the model bias. We found an increase in model bias at a few stations, primarily due to the coarse spatial resolution of the CMAQ model (12 km), which cannot resolve the peak of surface NO₂ concentrations in a small area. However, out of 230 stations, 153 exhibited a decrease in a bias of surface NO₂ concentrations with the posterior emissions. This enhanced model performance with the posterior NO_x emissions was confirmed by Fig. 4, showing the histogram of the daytime averaged difference between modeled surface NO₂ concentrations and AQS measurements. With the prior NO_x emissions, modeled surface NO₂ concentrations showed a largely left-skewed distribution with a median of -1.53 ppb. The posterior emissions constrained by the daily complete proxy NO₂ map exhibited better performance with a significantly reduced bias (0.26 ppb), demonstrating the advantage of using DL-estimated proxy surface NO₂ concentrations, which are more capable of informing ground-level air pollution over a broad area compared to either surface or satellite data.

Inversions with satellite data are often conducted at a monthly time resolution due to the high missing ratio, which requires accumulating data over a certain period to cover the entire study domain. Consequently, the monthly emissions adjustment ratio from top-down inversions can monotonically increase or decrease emissions each month, hindering our ability to ascertain the variability of NO_x caused by robust changes in both emission sources’ activity and meteorological conditions. In this regard, using daily complete proxy surface NO₂ can be of considerable help in quantifying NO_x emissions at a finer temporal resolution and even capturing sudden changes that occur within a month. Figure 5 shows the time series of daytime-averaged surface NO₂ concentrations simulated with prior and posterior NO_x emissions, collocated with AQS measurements over the CONUS in the summer of 2017, aligning with correlation and mean absolute percentage error (MAPE). Additionally, we plotted three areas (Box A to Box C, denoted in Fig. 3) with high daily variability. Over the CONUS, as shown in Fig. 5 (a), the model with prior emissions tended to consistently underestimate NO₂ concentrations with a flattened amplitude compared to the measurement data throughout the study period. Promisingly, the model performance with the posterior NO_x emissions significantly improved in terms of correspondence (0.84 to 0.92) and error reductions (0.48 ppb to 0.10 ppb). Furthermore, the model with the posterior emissions was able to capture a sudden peak reaching up to 5.52 ppb on August 2, which was beyond the range estimated during the study period (1.57 ± 0.31 ppb) (Fig. 5 (a)). In contrast, the model with prior emissions showed significantly lower NO₂ concentrations at 2.55 ppb. This improvement in model performance, achieved by incorporating the daily completed proxy surface NO₂ map, was also observed in areas with high day-to-day variability (Fig. 5 (b) to (d)).

This study focused on the synergy between DL and numerical modeling in quantifying NO_x emissions at a fine spatiotemporal resolution over the CONUS in the summer of 2017. To train the PCNN and DNN models for imputing the OMI TCDNO₂ and estimating daily complete proxy surface NO₂ map, respectively, we used modeled surface NO₂ concentrations with the prior NO_x emissions along with several predictor data. The approach enabled DL models to have a large number of training samples with no gaps, thereby improving the performance of the DL model. Subsequently, we estimated posterior NO_x emissions based on the final product of DL through numerical modeling, which showed improved performance in simulating surface NO₂ concentrations at a fine resolution, thereby enhancing our understanding of the chemical evolution. This approach could also be applied in estimating other atmospheric constituents, such as ammonia and sulfur dioxide, which often have many missing values in their satellite column measurements.

Promisingly, using the daily complete proxy surface NO₂ map estimated by using DL led to an improved model performance with an increased agreement with the AQS surface measurement. However, this study applied the IFDMB method as an inversion method that does not account for the uncertainty of DL, arising predominantly due to the mismatches between the training and test dataset, as well as errors in the dataset and parameters (Abdar et al., 2021). In recent years, several studies have been devoted to comparing various methods of uncertainty quantification in DL (e.g., Bayesian neural networks, concrete dropout, and deep ensembles learning techniques) and their performances (Abdar et al., 2021; Caldeira and Nord, 2020). Further investigations need to trace the accuracy and efficacy of DL associated with the application in the imputation of satellite data and estimation of surface NO₂ concentration. Another source of uncertainty may arise from the daily NO_x emission adjustment based on the overpass time of the OMI satellite (18 UTC). Factors such as emission sources (e.g., vehicles and power plants) and diurnal evolution of chemistry (involving NO_x, O₃, and hydroperoxyl radical), transport, and vertical mixing in the planetary boundary layer contribute to site-to-site variations in diurnal patterns. Therefore, high measurement accuracy is required to accurately represent diurnal variations of surface NO₂ concentrations, which warrants further application of DL and numerical modeling.

PCNN model for imputing OMI tropospheric NO₂ column images

Images captured by satellite instruments, such as the OMI, often contain missing values due to factors like cloud fraction and sensor failures. To address this issue and obtain complete images by imputing missing pixels in OMI TCDNO₂, we designed the PCNN model (Lops et al., 2023). The PCNN model consists of a U-net deep convolutional neural network (Ronneberger et al., 2015) that incorporates partial convolution layers (Liu et al., 2018) in addition to regular convolution layers (Krizhevsky et al., 2017).

Convolution neural networks (CNNs) process data by convolving the data at each layer across multiple image channels, assigning weights and biases at various aspects within the data, and differentiating between them. The partial convolution padding process in the PCNN model gradually reduces the influence of the missing data mask during each encoding phase. The partial convolution process considers only valid pixels, and normalization is adjusted for a fraction of missing data. To handle different mask features for each of the three channels of the dataset (representing the temporal dimension), we replaced the conventional 2D convolution layer with depthwise convolutions (Chollet, 2017) within the partial convolution process during the encoding phase of the model. The PCNN model was trained using TCDNO₂ simulated by the Community Multiscale Air Quality (CMAQ) modeling system version 5.2, incorporating OMI missing data masks. Once trained, the model performs spatiotemporal imputation of the original OMI datasets, resulting in a complete OMI TCDNO₂ that can be used by the DNN model for proxy NO₂ surface estimations. For more detailed information about the PCNN model used in this study, refer to Lops et al. (2023).

DNN model for estimating daily proxy surface NO₂ map

We employed the DNN model to estimate daily surface concentrations of NO₂ over the CONUS in the summer of 2017 at a spatial resolution of 10×10 km. The DNN model consists of input and output layers connecting to four hidden layers, each containing 70 neurons (Ghahremanloo et al., 2023b).

In our analysis, we utilized several predictor variables, including the OMI TCDNO₂, which was imputed by the PCNN model, as mentioned earlier. We also incorporated weight-averaged NO₂ concentrations (10×10 km), generated using the distance-inversed weighted average approach (Di et al., 2016) with data from the Environmental Protection Agency (EPA) Air Quality System (AQS), Other predictor variables included the Enhanced Vegetation Index (EVI) (5×5 km) from the Moderate Resolution Imaging Spectroradiometer (MODIS), various meteorological factors (0.125°×0.125°) such as air temperature, surface pressure, specific humidity, U/V components of wind speed, and longwave radiation flux downward from the North American Land Data Assimilation System (NLDAS), surface layer height (0.5°×0.625°) from the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2), modeled surface NO₂ concentrations (12 km) from the CMAQ simulation using a priori NO_x emissions, population density (8 km) from the NASA Socioeconomic Data and Applications Center (Center for International Earth Science Information Network, 2018), road density (8 km) (km per km²) from the Global Roads Inventory Project (Meijer et al., 2018), and percentage of urban space (8 km) from the 2016 National Land Cover Database (NLCD). We matched all predictor variables to the closest grid cells of the weight-averaged NO₂ layer. To assess multicollinearity among the predictor variables and determine feature importance, we employed the variance inflation factor (VIF) approach (Ghahremanloo et al., 2021b; Kock et al., 2012; RB Kline, 2015) and the SHapley Additive exPlanations (SHAP) feature importance (Ghahremanloo et al., 2021b).

WRF-CMAQ modeling setup

We updated NO_x emissions using the Weather Research and Forecasting (WRF)-CMAQ simulation to gain insights into the subsequent impacts on air chemistry. The CMAQ version 5.2, developed and released by the US Environmental Protection Agency (EPA), was utilized to simulate air quality variables, including surface NO₂ concentration, and to capture the relationship between NO_x emissions and NO₂ concentration. Meteorological input data was generated by WRF version 4.0, with the initial and boundary conditions obtained from the National Centers for Environmental Prediction (NCEP) North American Regional Reanalysis (NARR). The model simulations were conducted over the CONUS during the summer of 2017, using a horizontal grid spacing of 12 km. The emission data employed as a priori for NO_x emission inversion comprised the US EPA NEI 2017 (Eyth et al., 2016; Eyth and Vukovich, 2016) for anthropogenic emission, biogenic emission computed using the Biogenic Emission Inventory System (BEIS) version 3.61 incorporated within the SMOKE modeling system, biomass burning emissions based on the Fire Inventory from National Center for Atmospheric Research (FINN) version 1.5 (Wiedinmyer et al., 2014, 2011, 2006), and lightning emission produced by the in-line lightning NO (LNO) production module in the CMAQ (Allen et al., 2012; Kang et al., 2019) with data from the National Lightning Detection Network (NLDN). Considering the significant contribution of background sources in the free troposphere to tropospheric NO_x (Silvern et al., 2019), we implemented dynamic lateral boundary conditions by incorporating additional hemispheric modeling (HCMAQ) with seasonal scaling based on multiple satellites (e.g., OMI/Aura vertical O₃ profile, tropospheric NO₂ and HCHO, and MOPITT CO). The dynamic lateral boundary conditions, together with lightning NO_x emissions, effectively reduced the underestimation of NO₂ in the upper troposphere by 23.24% (Jung et al., 2022a). Further details regarding the modeling setup and input data preparation can be found in the earlier study by Jung et al. (2022a).

Iterative finite difference mass balance (IFDMB) method

The mass balance approach is utilized to estimate the posterior emissions based on the relationship between NO_x emissions and NO₂ concentrations near the surface. Additionally, a scaling factor β, introduced by Lamsal et al. (2011), is incorporated to reflect the sensitivity of changes in NO₂ concentrations to changes in NO_x emissions, and it helps reduce errors associated with the chemical process of NO_x (e.g., NO_x oxidation by OH and N₂O₅ hydrolysis). In this study, we updated NO_x emissions using the IFDMB method, following the approach described by Cooper et al. (2017), which iteratively adjusts the NO_x emissions until the normalized mean difference between the new emissions and the emissions from the previous iteration becomes less than 1%.

$${E}_{t}={E}_{a}\left(1+\frac{1}{\beta }\frac{{{\Omega }}_{o}-{{\Omega }}_{a}}{{{\Omega }}_{a}}\right)$$

where ${E}_{t}$ and ${E}_{a}$ represent the posteriori and priori emissions, respectively. Ωo represents the DL-estimated complete surface map, and Ω_a represents the simulated NO₂ concentrations. β is the scaling factor, which accounts for the sensitivity of NO₂ concentrations to changes in NO_x emissions, calculated by (∆Ω/Ω)/(∆E/E).

OMI tropospheric NO₂ column

The OMI instrument on the NASA Earth Observing System (EOS) Aura satellite was launched in July 2004. Since its launch, OMI has been providing continuous daily measurements of several air quality components (such as NO₂, sulfur dioxide (SO₂), bromine monoxide (BrO), HCHO, and aerosols) at a relatively high spatial resolution (up to 13 × 24 km² at the nadir) (Levelt et al., 2006). In this study, the tropospheric NO₂ columns used are the OMI operational retrieval products (Level 2 and version 3) distributed by the NASA Goddard Earth Sciences Data and Information Service Center (GES DISC) (Bucsela et al., 2013; Choi et al., 2020; Krotkov et al., 2017; Lamsal et al., 2014). However, similar to other satellite measurements, a significant amount of data is missing due to various factors such as cloud cover, cloud shadow, surface contamination, and erroneous surface reflectance caused by clouds and aerosol (Choi et al., 2016). Furthermore, the presence of row anomaly in the OMI data since 2008 has led to a significant reduction in the number of available data points and spatial coverage, impeding its further application (Torres et al., 2018).

Acknowledgments: This study was supported by the University of Houston High Priority study and NASA Aura Science Team Grant (NNH19ZDA001N-AURAST). Feng is supported by NASA grant #80HQTR21T0069 at Pacific Northwest National Laboratory (PNNL). PNNL is operated for the Department of Energy by Battelle Memorial Institute under contract DE-AC0676RL01830.

Data Availability Statement: The EPA AQS data (pre-generated) for meteorology and air quality measurement over the U.S. is available from https://aqs.epa.gov/aqsweb/airdata/download_files.html. The OMI/Aura Level 2 tropospheric NO2 column data product is downloaded from the NASA Goddard Earth Science (GES) Data and Information Services Center (DISC) (https://disc.gsfc.nasa.gov/datasets/OMNO2_003/summary?keywords=no2). The model outputs with prior and posterior NO_x emissions and DL-estimated NO₂ concentration can be downloaded from https://doi.org/10.5281/zenodo.8072167.

Author contributions: J.J performed simulations and wrote the paper. J.J. and Y.C. planned the research. M.G. and Y.L. performed deep learning estimation and provided input data. M.M. contributed to code development. S.F., S.P., W.J., and J.W.Y. provided discussions and comments on the paper.

Abdar, M., Pourpanah, F., Hussain, S., Rezazadegan, D., Liu, L., Ghavamzadeh, M., Fieguth, P., Cao, X., Khosravi, A., Acharya, U.R., Makarenkov, V., Nahavandi, S., 2021. A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Information Fusion 76, 243–297. https://doi.org/10.1016/j.inffus.2021.05.008
Allen, D.J., Pickering, K.E., Pinder, R.W., Henderson, B.H., Appel, K.W., Prados, A., 2012. Impact of lightning-NO on eastern United States photochemistry during the summer of 2006 as determined using the CMAQ model. Atmospheric Chemistry and Physics 12, 1737–1758. https://doi.org/10.5194/acp-12-1737-2012
Bucsela, E.J., Krotkov, N.A., Celarier, E.A., Lamsal, L.N., Swartz, W.H., Bhartia, P.K., Boersma, K.F., Veefkind, J.P., Gleason, J.F., Pickering, K.E., 2013. A new stratospheric and tropospheric NO2 retrieval algorithm for nadir-viewing satellite instruments: applications to OMI. Atmospheric Measurement Techniques 6, 2607–2626. https://doi.org/10.5194/amt-6-2607-2013
Caldeira, J., Nord, B., 2020. Deeply uncertain: comparing methods of uncertainty quantification in deep learning algorithms. Mach. Learn.: Sci. Technol. 2, 015002. https://doi.org/10.1088/2632-2153/aba6f3
Chai, T., Carmichael, G.R., Tang, Y., Sandu, A., Heckel, A., Richter, A., Burrows, J.P., 2009. Regional NOx emission inversion through a four-dimensional variational approach using SCIAMACHY tropospheric NO2 column observations. Atmospheric Environment 43, 5046–5055. https://doi.org/10.1016/j.atmosenv.2009.06.052
Choi, M., Kim, J., Lee, J., Kim, M., Park, Y.-J., Jeong, U., Kim, W., Hong, H., Holben, B., Eck, T.F., Song, C.H., Lim, J.-H., Song, C.-K., 2016. GOCI Yonsei Aerosol Retrieval (YAER) algorithm and validation during the DRAGON-NE Asia 2012 campaign. Atmospheric Measurement Techniques 9, 1377–1398. https://doi.org/10.5194/amt-9-1377-2016
Choi, S., Lamsal, L.N., Follette-Cook, M., Joiner, J., Krotkov, N.A., Swartz, W.H., Pickering, K.E., Loughner, C.P., Appel, W., Pfister, G., Saide, P.E., Cohen, R.C., Weinheimer, A.J., Herman, J.R., 2020. Assessment of NO2 observations during DISCOVER-AQ and KORUS-AQ field campaigns. Atmospheric Measurement Techniques 13, 2523–2546. https://doi.org/10.5194/amt-13-2523-2020
Chollet, F., 2017. Xception: Deep Learning with Depthwise Separable Convolutions. arXiv:1610.02357 [cs].
Chu, P., Yuchun, C., Kuninaka, A., 2005. Seasonal variability of the Yellow Sea/East China Sea surface fluxes and thermohaline structure. Advances in Atmospheric Sciences 22, 1–20. https://doi.org/10.1007/BF02930865
Cooper, M., Martin, R.V., Padmanabhan, A., Henze, D.K., 2017. Comparing mass balance and adjoint methods for inverse modeling of nitrogen dioxide columns for global nitrogen oxide emissions: Mass Balance and Adjoint NO _x Inversions. J. Geophys. Res. Atmos. 122, 4718–4734. https://doi.org/10.1002/2016JD025985
Di, Q., Kloog, I., Koutrakis, P., Lyapustin, A., Wang, Y., Schwartz, J., 2016. Assessing PM _2.5 Exposures with High Spatiotemporal Resolution across the Continental United States. Environ. Sci. Technol. 50, 4712–4721. https://doi.org/10.1021/acs.est.5b06121
Di, Q., Rowland, S., Koutrakis, P., Schwartz, J., 2017. A hybrid model for spatially and temporally resolved ozone exposures in the continental United States. Journal of the Air & Waste Management Association 67, 39–52. https://doi.org/10.1080/10962247.2016.1200159
Eyth, A., Vukovich, J., 2016. Technical Support Document (TSD) preparation of emissions inventories for the version 6.3, 2011 emissions modeling platform. US Environmental Protection Agency, Office of Air Quality Planning and Standards.
Eyth, A., Vukovich, J., Farkas, C., Strum, M., 2016. Technical Support Document (TSD): Preparation of Emissions Inventories for the Version 7.1 - 2016 North American Emissions Modeling Platform.
Ghahremanloo, M., Choi, Y., Lops, Y., 2023a. Deep learning mapping of surface MDA8 ozone: The impact of predictor variables on ozone levels over the contiguous United States. Environmental Pollution 326, 121508. https://doi.org/10.1016/j.envpol.2023.121508
Ghahremanloo, M., Choi, Y., Sayeed, A., Salman, A.K., Pan, S., Amani, M., 2021a. Estimating daily high-resolution PM2.5 concentrations over Texas: Machine Learning approach. Atmospheric Environment 247, 118209. https://doi.org/10.1016/j.atmosenv.2021.118209
Ghahremanloo, M., Lops, Y., Choi, Y., Jung, J., Mousavinezhad, S., Hammond, D., 2022. A comprehensive study of the COVID-19 impact on PM2.5 levels over the contiguous United States: A deep learning approach. Atmospheric Environment 272, 118944. https://doi.org/10.1016/j.atmosenv.2022.118944
Ghahremanloo, M., Lops, Y., Choi, Y., Mousavinezhad, S., Jung, J., 2023b. A coupled deep learning model for estimating surface NO2 levels from remote sensing data: 15-year study over the contiguous United States. Journal of Geophysical Research: Atmospheres. https://doi.org/10.1029/2022JD037010
Ghahremanloo, M., Lops, Y., Choi, Y., Yeganeh, B., 2021b. Deep Learning Estimation of Daily Ground‐Level NO ₂ Concentrations From Remote Sensing Data. Geophys Res Atmos 126. https://doi.org/10.1029/2021JD034925
Goldberg, D.L., Saide, P.E., Lamsal, L.N., de Foy, B., Lu, Z., Woo, J.-H., Kim, Y., Kim, J., Gao, M., Carmichael, G., Streets, D.G., 2019. A top-down assessment using OMI NO2 suggests an underestimate in the NOx emissions inventory in Seoul, South Korea, during KORUS-AQ. Atmospheric Chemistry and Physics 19, 1801–1818. https://doi.org/10.5194/acp-19-1801-2019
Jiang, Z., McDonald, B.C., Worden, H., Worden, J.R., Miyazaki, K., Qu, Z., Henze, D.K., Jones, D.B.A., Arellano, A.F., Fischer, E.V., Zhu, L., Boersma, K.F., 2018. Unexpected slowdown of US pollutant emission reduction in the past decade. Proceedings of the National Academy of Sciences 115, 5099–5104. https://doi.org/10.1073/pnas.1801191115
Jung, J., Choi, Y., Mousavinezhad, S., Kang, D., Park, J., Pouyaei, A., Ghahremanloo, M., Momeni, M., Kim, H., 2022a. Changes in the ozone chemical regime over the contiguous United States inferred by the inversion of NOx and VOC emissions using satellite observation. Atmospheric Research 270, 106076. https://doi.org/10.1016/j.atmosres.2022.106076
Jung, J., Choi, Y., Souri, A.H., Mousavinezhad, S., Sayeed, A., Lee, K., 2022b. The Impact of Springtime‐Transported Air Pollutants on Local Air Quality With Satellite‐Constrained NO _x Emission Adjustments Over East Asia. JGR Atmospheres 127. https://doi.org/10.1029/2021JD035251
Kang, D., Foley, K.M., Mathur, R., Roselle, S.J., Pickering, K.E., Allen, D.J., 2019. Simulating lightning NO production in CMAQv5.2: performance evaluations. Geosci. Model Dev. 12, 4409–4424. https://doi.org/10.5194/gmd-12-4409-2019
Kock, N., Lynn, G., Stevens Institute of Technology, 2012. Lateral Collinearity and Misleading Results in Variance-Based SEM: An Illustration and Recommendations. JAIS 13, 546–580. https://doi.org/10.17705/1jais.00302
Konovalov, I.B., Beekmann, M., Richter, A., Burrows, J.P., 2006. Inverse modelling of the spatial distribution of NOx emissions on a continental scale using satellite data. Atmospheric Chemistry and Physics 25.
Krizhevsky, A., Sutskever, I., Hinton, G.E., 2017. ImageNet classification with deep convolutional neural networks. Commun. ACM 60, 84–90. https://doi.org/10.1145/3065386
Krotkov, N.A., Lamsal, L.N., Celarier, E.A., Swartz, W.H., Marchenko, S.V., Bucsela, E.J., Chan, K.L., Wenig, M., Zara, M., 2017. The version 3 OMI NOs standard product. Atmospheric Measurement Techniques 10, 3133–3149. https://doi.org/10.5194/amt-10-3133-2017
Kurokawa, J., Yumimoto, K., Uno, I., Ohara, T., 2009. Adjoint inverse modeling of NOx emissions over eastern China using satellite observations of NO2 vertical column densities. Atmospheric Environment 43, 1878–1887. https://doi.org/10.1016/j.atmosenv.2008.12.030
Lamsal, L.N., Krotkov, N.A., Celarier, E.A., Swartz, W.H., Pickering, K.E., Bucsela, E.J., Gleason, J.F., Martin, R.V., Philip, S., Irie, H., Cede, A., Herman, J., Weinheimer, A., Szykman, J.J., Knepp, T.N., 2014. Evaluation of OMI operational standard NO2 column retrievals using in situ and surface-based NO2 observations. Atmospheric Chemistry and Physics 14, 11587–11609. https://doi.org/10.5194/acp-14-11587-2014
Lamsal, L.N., Martin, R.V., Padmanabhan, A., van Donkelaar, A., Zhang, Q., Sioris, C.E., Chance, K., Kurosu, T.P., Newchurch, M.J., 2011. Application of satellite observations for timely updates to global anthropogenic NO _x emission inventories: UPDATING NO _x EMISSION INVENTORIES. Geophys. Res. Lett. 38, n/a-n/a. https://doi.org/10.1029/2010GL046476
Lei, H., Wang, J.X.L., 2014. Sensitivities of NO<sub>x</sub> transformation and the effects on surface ozone and nitrate. Atmos. Chem. Phys. 14, 1385–1396. https://doi.org/10.5194/acp-14-1385-2014
Levelt, P.F., van den Oord, G.H.J., Dobber, M.R., Malkki, A., Huib Visser, Johan de Vries, Stammes, P., Lundell, J.O.V., Saari, H., 2006. The ozone monitoring instrument. IEEE Transactions on Geoscience and Remote Sensing 44, 1093–1101. https://doi.org/10.1109/TGRS.2006.872333
Li, L., Wu, J., 2021. Spatiotemporal estimation of satellite-borne and ground-level NO2 using full residual deep networks. Remote Sensing of Environment 254, 112257. https://doi.org/10.1016/j.rse.2020.112257
Liu, G., Reda, F.A., Shih, K.J., Wang, T.-C., Tao, A., Catanzaro, B., 2018. Image Inpainting for Irregular Holes Using Partial Convolutions. arXiv:1804.07723 [cs].
Liu, J., Weng, F., Li, Z., 2019a. Satellite-based PM2.5 estimation directly from reflectance at the top of the atmosphere using a machine learning algorithm. Atmospheric Environment 208, 113–122. https://doi.org/10.1016/j.atmosenv.2019.04.002
Liu, J., Weng, F., Li, Z., Cribb, M.C., 2019b. Hourly PM2.5 Estimates from a Geostationary Satellite Based on an Ensemble Learning Algorithm and Their Spatiotemporal Patterns over Central East China. Remote Sensing 11, 2120. https://doi.org/10.3390/rs11182120
Liu, R., Ma, Z., Liu, Y., Shao, Y., Zhao, W., Bi, J., 2020. Spatiotemporal distributions of surface ozone levels in China from 2005 to 2017: A machine learning approach. Environment International 142, 105823. https://doi.org/10.1016/j.envint.2020.105823
Lops, Y., Ghahremanloo, M., Pouyaei, A., Choi, Y., Jung, J., Mousavinezhad, S., Salman, A.K., Hammond, D., 2023. Spatiotemporal estimation of TROPOMI NO2 column with depthwise partial convolutional neural network. Neural Comput & Applic 35, 15667–15678. https://doi.org/10.1007/s00521-023-08558-1
Lops, Y., Pouyaei, A., Choi, Y., Jung, J., Salman, A.K., Sayeed, A., 2021. Application of a Partial Convolutional Neural Network for Estimating Geostationary Aerosol Optical Depth Data. Geophys Res Lett 48. https://doi.org/10.1029/2021GL093096
Martin, R.V., 2003. Global inventory of nitrogen oxide emissions constrained by space-based observations of NO ₂ columns. J. Geophys. Res. 108, 4537. https://doi.org/10.1029/2003JD003453
Meijer, J.R., Huijbregts, M.A.J., Schotten, K.C.G.J., Schipper, A.M., 2018. Global patterns of current and future road infrastructure. Environ. Res. Lett. 13, 064006. https://doi.org/10.1088/1748-9326/aabd42
Mun, J., Choi, Y., Jeon, W., Lee, H.W., Kim, C.-H., Park, S.-Y., Bak, J., Jung, J., Oh, I., Park, J., Kim, D., 2023. Assessing mass balance-based inverse modeling methods via a pseudo-observation test to constrain NOx emissions over South Korea. Atmospheric Environment 292, 119429. https://doi.org/10.1016/j.atmosenv.2022.119429
Nabavi, S.O., Nölscher, A.C., Samimi, C., Thomas, C., Haimberger, L., Lüers, J., Held, A., 2021. Site-scale modeling of surface ozone in Northern Bavaria using machine learning algorithms, regional dynamic models, and a hybrid model. Environmental Pollution 268, 115736. https://doi.org/10.1016/j.envpol.2020.115736
Park, Y., Kwon, B., Heo, J., Hu, X., Liu, Y., Moon, T., 2020. Estimating PM2.5 concentration of the conterminous United States via interpretable convolutional neural networks. Environmental Pollution 256, 113395. https://doi.org/10.1016/j.envpol.2019.113395
RB Kline, 2015. Principles and practice of structural equation modeling. Guilford publications.
Ren, X., Mi, Z., Georgopoulos, P.G., 2020. Comparison of Machine Learning and Land Use Regression for fine scale spatiotemporal estimation of ambient air pollution: Modeling ozone concentrations across the contiguous United States. Environment International 142, 105827. https://doi.org/10.1016/j.envint.2020.105827
Requia, W.J., Di, Q., Silvern, R., Kelly, J.T., Koutrakis, P., Mickley, L.J., Sulprizio, M.P., Amini, H., Shi, L., Schwartz, J., 2020. An Ensemble Learning Approach for Estimating High Spatiotemporal Resolution of Ground-Level Ozone in the Contiguous United States. Environ. Sci. Technol. 54, 11037–11047. https://doi.org/10.1021/acs.est.0c01791
Ronneberger, O., Fischer, P., Brox, T., 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv:1505.04597 [cs].
Schneider, P., van der A, R.J., 2012. A global single-sensor analysis of 2002-2011 tropospheric nitrogen dioxide trends observed from space: GLOBAL NO ₂ TRENDS OBSERVED FROM SPACE. J. Geophys. Res. 117, n/a-n/a. https://doi.org/10.1029/2012JD017571
Silvern, R.F., Jacob, D.J., Mickley, L.J., Sulprizio, M.P., Travis, K.R., Marais, E.A., Cohen, R.C., Laughner, J.L., Choi, S., Joiner, J., Lamsal, L.N., 2019. Using satellite observations of tropospheric NO ₂ columns to infer long-term trends in US NO _x emissions: the importance of accounting for the free tropospheric NO ₂ background. Atmospheric Chemistry and Physics 19, 8863–8878. https://doi.org/10.5194/acp-19-8863-2019
Souri, A.H., Choi, Y., Jeon, W., Li, X., Pan, S., Diao, L., Westenbarger, D.A., 2016. Constraining NOx emissions using satellite NO2 measurements during 2013 DISCOVER-AQ Texas campaign. Atmospheric Environment 131, 371–381. https://doi.org/10.1016/j.atmosenv.2016.02.020
Torres, O., Bhartia, P.K., Jethva, H., Ahn, C., 2018. Impact of the ozone monitoring instrument row anomaly on the long-term record of aerosol products. Atmospheric Measurement Techniques 11, 2701–2715. https://doi.org/10.5194/amt-11-2701-2018
Wei, J., Huang, W., Li, Z., Xue, W., Peng, Y., Sun, L., Cribb, M., 2019. Estimating 1-km-resolution PM2.5 concentrations across China using the space-time random forest approach. Remote Sensing of Environment 231, 111221. https://doi.org/10.1016/j.rse.2019.111221
Wei, J., Li, Z., Cribb, M., Huang, W., Xue, W., Sun, L., Guo, J., Peng, Y., Li, J., Lyapustin, A., Liu, L., Wu, H., Song, Y., 2020. Improved 1 km resolution PM<sub>2.5</sub> estimates across China using enhanced space–time extremely randomized trees. Atmos. Chem. Phys. 20, 3273–3289. https://doi.org/10.5194/acp-20-3273-2020
Wiedinmyer, C., Akagi, S.K., Yokelson, R.J., Emmons, L.K., Al-Saadi, J.A., Orlando, J.J., Soja, A.J., 2011. The Fire INventory from NCAR (FINN): a high resolution global model to estimate the emissions from open burning. Geoscientific Model Development 4, 625–641. https://doi.org/10.5194/gmd-4-625-2011
Wiedinmyer, C., Quayle, B., Geron, C., Belote, A., McKenzie, D., Zhang, X., O’Neill, S., Wynne, K.K., 2006. Estimating emissions from fires in North America for air quality modeling. Atmospheric Environment 40, 3419–3432. https://doi.org/10.1016/j.atmosenv.2006.02.010
Wiedinmyer, C., Yokelson, R.J., Gullett, B.K., 2014. Global Emissions of Trace Gases, Particulate Matter, and Hazardous Air Pollutants from Open Burning of Domestic Waste. Environmental Science & Technology 48, 9523–9530. https://doi.org/10.1021/es502250z
Xu, J.-W., Martin, R.V., van Donkelaar, A., Kim, J., Choi, M., Zhang, Q., Geng, G., Liu, Y., Ma, Z., Huang, L., Wang, Y., Chen, H., Che, H., Lin, P., Lin, N., 2015. Estimating ground-level PM2.5 in eastern China using aerosol optical depth determined from the GOCI satellite instrument. Atmospheric Chemistry and Physics 15, 13133–13144. https://doi.org/10.5194/acp-15-13133-2015
Yeganeh, B., Hewson, M.G., Clifford, S., Tavassoli, A., Knibbs, L.D., Morawska, L., 2018. Estimating the spatiotemporal variation of NO2 concentration using an adaptive neuro-fuzzy inference system. Environmental Modelling & Software 100, 222–235. https://doi.org/10.1016/j.envsoft.2017.11.031

No competing interests reported.

NOxDLsupplementaryfile.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

The synergy between deep learning and numerical modeling in estimating NO_x emissions at a fine spatiotemporal resolution

Status:

Version 1

Abstract

Figures

Introduction

Results

Estimation of Daily Complete Proxy Surface NO₂ Map and its Validation

Inversion of NO_x Emissions Using Daily Complete Proxy Surface NO₂ Map

Discussion

Method

PCNN model for imputing OMI tropospheric NO₂ column images

DNN model for estimating daily proxy surface NO₂ map

WRF-CMAQ modeling setup

Iterative finite difference mass balance (IFDMB) method

OMI tropospheric NO₂ column

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1

The synergy between deep learning and numerical modeling in estimating NOx emissions at a fine spatiotemporal resolution

Status:

Version 1

Abstract

Figures

Introduction

Results

Estimation of Daily Complete Proxy Surface NO2 Map and its Validation

Inversion of NOx Emissions Using Daily Complete Proxy Surface NO2 Map

Discussion

Method

PCNN model for imputing OMI tropospheric NO2 column images

DNN model for estimating daily proxy surface NO2 map

WRF-CMAQ modeling setup

Iterative finite difference mass balance (IFDMB) method

OMI tropospheric NO2 column

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1

The synergy between deep learning and numerical modeling in estimating NO_x emissions at a fine spatiotemporal resolution

Estimation of Daily Complete Proxy Surface NO₂ Map and its Validation

Inversion of NO_x Emissions Using Daily Complete Proxy Surface NO₂ Map

PCNN model for imputing OMI tropospheric NO₂ column images

DNN model for estimating daily proxy surface NO₂ map

OMI tropospheric NO₂ column