IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.

60, 2022 4101113

Prediction of Vertical Profile of NO2 Using Deep

Multimodal Fusion Network Based on the
Ground-Based 3-D Remote Sensing
Shulin Zhang , Bo Li, Lei Liu , Qihou Hu, Haoran Liu, Rui Zheng , Yizhi Zhu,
Ting Liu, Mingzhai Sun , and Cheng Liu

Abstract— The vertical distribution profiles of NO2 are essen- of the major pollutants in the atmosphere, is toxic to human
tial for understanding the mechanisms, detecting near-surface health and causes severe damage to the immune and respira-
emissions, and tracking pollutant transportation at high altitude. tory systems. The long-term exposure to NO2 can increase the
However, most of the published NO2 studies are based on the
surface 2-D measurements. The ground-based 3-D remote-sensing risk of all-cause mortality [1], even hurt the plant [2]. More-
stations were recently built to measure vertical distribution over, NO2 is the precursor of secondary aerosols nitrate and
profiles of NO2 . However, the stations were spatially sparse due has a significant impact on the formation and destruction of
to the high cost and could not make the measurements without other air pollutants. It catalyzes the production of ozone (O3 )
sunlight. In this study, we first developed a multimodel fusion in the troposphere and the removal of O3 in the stratospheric.
network (MF-net) based on the sparse vertical observations from
the Jing-Jin-Ji region. We achieved the 3-D profile prediction Therefore, NO2 plays a vital role in both tropospheric and
of NO2 in the range of 39.005–41.405N and 115.005–117.905E stratospheric chemistry. Thus, the prevention and control of
with 24-h coverage. The MF-net significantly surpassed the NO2 pollution are essential.
conventional WRF-CHEM model and provided a more accurate Accurate measurement of the NO2 concentration is the
evaluation of the NO2 transmission between Beijing and the premise to prevent and control of NO2 pollution, as well as
neighboring cities. Besides, the MF-net covers the monitoring
of NO2 to the whole study area and extends the monitoring time to study the pollution mechanism. Most of the published NO2
to the entire day (24 h), making it serviceable for continuous studies are based on the surface measurements of NO2 , such
spatial-temporal estimation of NO2 and its transmission in as the historical observations of the China National Envi-
pollution events. The MF-net provides more robust data support ronmental Monitoring Center (CNEMC). However, surface
to formulate reasonable and effective pollution prevention and NO2 measurement is not sufficient to fully characterize the
control measures.
pollution, since the pollution is caused not only by the local
Index Terms— 3-D prediction of NO2 , deep learning neural anthropogenic emissions (industrial emissions, power genera-
network, multiaxis differential optical absorption spectroscopy tion, and traffic emissions), but also remote transportation [3]
(MAX-DOAS), multimodal information fusion, remote sensing.
that happens at high altitudes.
To this end, a type of passive remote-sensing measure-
I. I NTRODUCTION ment system called multiaxis differential optical absorption
spectroscopy (MAX-DOAS) has been proposed [4]. And it
I N RECENT years, people have increasingly been con-
cerned about air pollution with the development of urban
industry and transportation. Nitrogen dioxide (NO2 ), as one
has been widely used in the past decade to measure the
vertical structure of tropospheric aerosol [5]–[7] and trace
gases [8]–[10] by applying the differential optical absorption
Manuscript received September 23, 2020; revised January 3, 2021; accepted spectroscopy (DOAS) method to the observations in several
February 4, 2021. Date of publication March 22, 2021; date of current
version December 9, 2021. This work was supported in part by the National different viewing directions. The spatial distribution data of
Natural Science Foundation of China under Grant 41722501, in part by the the observations are retrieved from the molecular absorption
National Key Research and Development Program of China under Grant in the ultraviolet (UV) and visible (VIS) wavelength bands
2018YFC0213104, in part by the National High-Resolution Earth Observation
Project of China under Grant 05-Y30B01-9001-19/20-3, in part by the measured from the scattered sunlight spectra. In this study,
Key Research Program of Frontier Sciences, CAS, under Grant ZDBS- the dark current and offset spectra were used to correct
LY-DQC008, and in part by the National Key Research and Development measurement spectra before the spectra analysis. And we
Program of China under Grant 2017YFC0210002. (Corresponding authors:
Mingzhai Sun; Cheng Liu.) analyzed the MAX-DOAS spectra by the QDOAS spectral
Shulin Zhang, Lei Liu, Rui Zheng, Mingzhai Sun, and Cheng Liu are fitting software suite developed by BIRA-IASB (http://uv-
with the Department of Precision Machinery and Instrumentation, Univer- vis.aeronomie.be/software/QDOAS/). The wavelength was cal-
sity of Science and Technology of China, Hefei 230026, China (e-mail:
[email protected]; [email protected]). ibrated with a high-resolution solar reference spectrum [11].
Bo Li, Yizhi Zhu, and Ting Liu are with the School of Earth and Space The DOAS fit included several trace gas absorption cross
Science, University of Science and Technology of China, Hefei 230026, sections [12]–[15], the Ring spectrum, a Frauenhofer reference
China.
Qihou Hu is with the Key Laboratory of Environmental Optics and spectrum, and a low-order polynomial. The details can be
Technology, Anhui Institute of Optics and Fine Mechanics, Chinese Academy found in our previous work [3], [16]. Currently, the available
of Sciences, Hefei 230031, China. column densities obtained from MAX-DOAS includes the
Haoran Liu is with the Institute of Physical Science and Information
Technology, Anhui University, Hefei 230601, China. atmospheric trace gases (e.g., nitrogen oxides (NOx) [3],
Digital Object Identifier 10.1109/TGRS.2021.3061476 [17], sulfur dioxide (SO2 ) [18], formaldehyde (HCHO) [19],
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
4101113 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

1) We achieved 3-D prediction of the vertical profile of

NO2 distribution using the proposed neural network
(MF-net) trained on the observations. The MF-net was
able to cover the entire study area and extend the mon-
itoring time from daytime to the whole day, including
night.
2) With the MF-net, we were able to evaluate the 3-D NO2
transmission between Beijing and the neighboring cities
more accurately from October 2016 to November 2019.
3) We applied the MF-net to study the dynamics of a
pollution event with a much higher spatial-temporal
continuity, which help understand the mechanism of
NO2 pollution in Beijing and formulating air quality
assurance policies.
This article was organized as follows. In Section II,
we introduced related researches about air pollutant prediction.
In Section III, we described the details about the data we
used in our study. The proposed 3-D prediction approach of
MF-net was also introduced in this section, as well as the
experimental process of the training and testing. Next, we first
compared the performance of MF-net and WRF-CHEM in
four experiments in Section IV-A. Then we demonstrated the
application value of MF-net by applying the algorithm in three
Fig. 1. Location of the MAX-DOAS instruments and the CNEMC sites.
practical applications in Section IV-B. The discussion and
The red dots represent the four MAX-DOAS stations with longer term conclusion of this work were in Sections V and VI.
measurements, while the greens marked the four MAX-DOAS sites containing
shorter term observations, and the blue triangles are 34 Beijing CNEMC sites.
The green line represented the grid of WRF-CHEM with a spatial resolution II. R ELATED W ORK
of 20 × 20 km.
Current 3-D prediction of air pollution can be grouped into
two categories: simulation models and neural network-based
Oxygen dimer (O4 ) [20], [21], Nitrous acid (HONO) [9], Gly- models. To simulate both meteorological component and
oxal [22]), and tropospheric column densities, surface volume chemical processes, Byun [25] proposed a computationally
mixing ratios (VMRs), vertical profiles of aerosol extinction, attractive approach [the community model for air quality
and trace gas mixing ratios. The MAX-DOAS observations (CMAQ)] for it produced many chemical transport simulations
have been performed on a range of flexible applications, such with only a single meteorological data set in the retrospective
as ground-based stations [16] and mobile platforms (e.g., offline chemical transport simulations. To avoid failing in
cars [23], ships [19], and aircraft [24]). However, the obser- capturing the atmospheric processes with a less time scale
vation range of ground-based stations is limited in space and than the meteorological model, typically available once or
time. It can only observe the vertical profile of NO2 at the twice per hour, the weather research and forecasting (WRF)
observation site’s location during the day and is interrupted at model (http://www.mmm.ucar.edu/wrf/users/document.html)
night. To cover the entire urban area needs a large number of was developed with the applicability on many different scales
observation sites. However, it is impracticable to build a large for the dynamic cores and different choices for physical
amount of MAX-DOAS sites due to the high cost of building parameterizations. Bacon et al. [26] applied the unstructured
and maintenance. grid to implement a continuously varying horizontal and
In this study, we collected the historical data from four vertical grid resolution. And Satoh [27] finished a similar
MAX-DOAS stations in Beijing-Tianjin-Hebei (Jing-Jin-Ji) work in the Japanese numerical weather prediction model.
region (Fig. 1), which was especially severely polluted due Grell et al. [28], [29] implemented various chemical modules
to its intensive use of fossil fuel for dense population’s energy into the WRF model in the study MM5/Chem and WRF-
and transportation demand [7]. We proposed a NO2 3-D CHEM. In the WRF-CHEM model, both the meteorology and
prediction multimodal fusion network (MF-net) that could chemistry were fully coupled “online,” with higher temporal
fully use the rich information from different types of data, and spatial resolution.
including the simulation results of NO2 and the meteorological Deep learning methods have been widely used in solving
parameters from WRF-CHEM. After training with the vertical many complex problems, including atmospheric prediction.
profile of NO2 from MAX-DOAS measurement, the network Zheng et al. [30] proposed a semisupervised deep learning
shows superior performance than WRF-CHEM simulation in approach to infer the real-time and fine-grained air quality
all experimental results. The main contributions of the study information based on the air quality data and a variety of data
are as follows. sources (such as meteorology, traffic flow, human mobility,
ZHANG et al.: PREDICTION OF VERTICAL PROFILE OF NO2 USING DEEP MULTIMODAL FUSION NETWORK 4101113

the structure of road networks, and point of interests (POI). TABLE I

Zheng et al. [31] also developed a predictive model pro- D ETAILS OF THE MAX-DOAS-L DATA S ET
viding 48-h fine-grained air quality forecasts for four major
Chinese cities every hour. Qi et al. [32] proposed a deep
learning network called deep air learning (DAL) to solve
the interpolation, prediction, and feature analysis problems
in one model. To estimate P M2.5 , deep belief network and
CNN-LSTM were applied with considering spatiotemporally
correlation [33], [34]. Cabaneros et al. [35] and Valput TABLE II
et al. [36] achieved forecasting NO2 concentration by neural D ETAILS OF THE MAX-DOAS-S DATA S ET
networks.
Currently, there is no deep learning research based on the
sparse MAX-DOAS observations for several reasons. First,
the observation data are minimal because there are very few
stations, and most stations are constructed in recent years.
Besides, the MAX-DOAS instruments can only measure NO2
during the daytime, not at night. Second, the unfavorable
weather conditions interrupt the measurement process inter-
mittently and cause substantial measurement errors or missing
data. Third, the NO2 prediction is very challenging since it
involves both meteorological factors (such as wind speed and vertical profiles from the MAX-DOAS instruments were
direction, turbulence, radiation, clouds, and precipitation) and similar to our previous work [3], [16]. We separated the
chemical processes (such as deposition and transformations), observations of MAX-DOAS stations into two data sets by
as well as transport contribution. Fourth, the stations are so whether the period of historical measurement was more than
sparse that it makes the interpolation very difficult for most one year. The MAX-DOAS-L data set described in Table I
of the grid without historical data. consisted of four stations with longer term MAX-DOAS
Evident from the situations above, there are currently no observations while specifying the time range of one to three
deep-learning-based studies on 3-D prediction of NO2 using years. It included Gucheng (GC), Guokeda (GKD), Nancheng
MAX-DOAS observations. To this end, we proposed a multi- (NC), and Qikeyuan (QKY), with 4672, 512, 3458, and
modal fusion network (MF-net), which entirely uses the verti- 7076 samples, respectively. The MAX-DOAS-S (Table II)
cal profile information of NO2 from MAX-DOAS. The MF-net were acquired from stations at DQS, Wangdu (WD), Xianghe
fused the multimodal data, including meteorological data and (XH), and Luancheng (LC) in a shorter period. The longitude,
vertical concentration of NO2 from WRF-CHEM simulation, latitude, and the observation period of each MAX-DOAS
and achieved the 3-D prediction of NO2 at any grid point station were listed in Tables I and II. In this study, MAX-
with or without historical data. The stereoscopic monitoring DOAS-L was using for training and cross-validation, whereas
of NO2 in an area provides valuable information for locating MAX-DOAS-S was only used for validation.
the sources of local emission and transmission. To compare 2) Meteorological Data: The 3-D distribution of the tem-
the performance of the MF-net and the WRF-CHEM model, perature, pressure, longitude wind speed, latitude wind speed,
we systemically evaluated both methods on MAX-DOAS vertical wind speed, and the atmospheric pollutant of NO2 for
measurements and other testing data sets, such as the 2-D the input of the MF-net was simulated using the WRF-CHEM
NO2 surface measurement from the CNEMC sites. All exper- model [3] with version 3.7.1. The model domain was con-
imental results show that the MF-net significantly outperforms figured to cover the Jing-Jin-Ji and its surrounding regions
WRF-CHEM in NO2 prediction. with a grid resolution of 20 × 20 km and 44 vertical layers.
To initialize the boundary conditions of meteorological fields
III. M ETHODOLOGY for the WRF-CHEM simulation, we adopted the 6-h final
operational global analysis data generated from the National
A. Data Sets Center for Environmental Prediction (NCEP) with a spatial
In this section, we introduce all the data sets we applied in resolution of 1◦ × 1◦ . The inputs of the emission inventory
the study. were obtained from the Multiresolution Emission Inventory
1) Data Sets of the MAX-DOAS Observations: Our for China (MEIC, http://www.meicmodel.org/) [37], [38]. The
MAX-DOAS instruments provided the measurements and simulations were conducted month by month and the outputs
errors of NO2 concentration at daytime with a measure- from previous runs were used as the initial conditions of the
ment intermittent of 15 min. The observations contained following run. To minimize the influence of initial conditions,
16 vertical altitudes, from the surface layer to 3 km with the first five days of each month were simulated and discarded
an interval of 200 m. However, due to the high miss rate as a spin-up period. Details of model configuration options can
and large measurement errors above 1 km, only the mea- be found in our previous studies [1], [39].
surements of six altitude layers within 0∼1 km are used 3) Surface Observations: The surface NO2 observa-
in our study. The details of the way to obtain the NO2 tion data, which we used to validate the MF-net, were
4101113 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Fig. 2. Overview structure of the MF-net. Each color patch (second column) corresponded to one of the six types of MF-net inputs, denoting with [Batch
size, 4, 4, 44], showing that the patch size was 4 × 4 with 44 vertical layers. The batch size was the number of samples used to update the MF-net weights
in each iteration. Each gray box (third column) represented a feature extraction module and had the same structure consisting of two convolutional layers
followed by max pooling operation. The kernel size and output channels of the convolutional layer were provided inside of the blue boxes. The notation of
“pool, /2" represented a pooling layer with a stride of 2 × 2. The number in the magenta boxes marked the node number of the fully connection layer.

collected by 34 CNEMC stations located in Beijing vertical altitudes. As the input of MF-net is based on the
(http://beijingair.sinaapp.com/). The concentration of the sur- simulation of WRF-CHEM, the spatial resolution of MF-net
face air pollutants including air quality index (AQI), P M2.5 , is 20 × 20 km and the temporal resolution is 1 h.
NO2 , and S O2 was measured every hour between 2018 and
2019. C. Training and Testing MF-Net
4) Satellite Observations: The tropospheric vertical column In this study, to improve the stability of the algorithm,
densities (VCDs) of NO2 were retrieved from the TRO- we removed the observations of MAX-DOAS with large
POspheric Monitoring Instrument (TROPOMI) instrument measurement errors, including the root mean square (RMS)
based on the algorithm of Zhang et al. [40], [41]. In our during the DOAS fit was larger than 0.1 or the corresponding
previous work [41], we regridded the tropospheric NO2 VCDs aerosol optical depth was larger than 3. Then, we interpolated
per orbit to the resolution of 0.1◦ × 0.1◦ over eastern China the cleaned MAX-DOAS observations to the integer hour.
using a novel P-Spline method [42]. There was only one Specifically, the average of the observations during the half
measurement result per day in Beijing and its surrounding hour before and after each hour is used as the measurement
areas, and the corresponding observation time is 3:30 (UTC). result at the integer hour.
We applied the fourfold cross-validation on the MAX-
B. MF-Net Architecture DOAS-L data set. Specifically, the observations were split into
The architecture of the MF-net is shown in Fig. 2. The four subsets by site. To ensure complete independence of the
MF-net started with six feature extraction modules with subsets data, observations from the same site were assigned
the same architecture consisted of two convolutional layers into the same subsets. Then, we performed four experiments to
followed by max pooling operation. It ended by two fully ensure that each subset was used as a test set and the remaining
connection layers fused the features obtained from the six subsets as the training set.
feature extraction modules and H . H represented a 1-D To further test the performance of the MF-net, we first
time vector, which can be expressed as (t1 , t2 , t3 , . . . , t24 ) trained the MF-net with all of the observations in the MAX-
where ti equals 1 when the hour of the sample date was i , DOAS-L data set. Then we evaluated the trained MF-net
otherwise 0. As shown in Fig. 2, every input of MF-net was by testing on the other three data sets: the MAX-DOAS-S,
a 4 × 4 patch with 44 layers, which was cropped from the CNEMC, and the satellite data set, respectively. When testing
meteorological data (such as temperature, pressure, and the on the observations of 34 CNEMC stations, we applied two
wind speed of latitude, longitude, and vertical) and the vertical evaluation conditions. In the first experiment, we evaluated
NO2 distribution from the WRF-CHEM simulation centering the surface prediction of MF-net on the CNEMC observa-
at the target point. The predictions of MF-net were the vertical tions that were most accurately monitored from 21:00 to
distribution of NO2 concentrations at six attitudes, from the 11:00 next day of UTC hour. For the second experiment,
surface to 1 km with an interval of 200 m, at the target points. we analyzed the performance of MF-net for whole day
The loss function of the MF-net is defined as observations.
The MF-net was implemented with TensorFlow using
L(Y, Ŷ ) = log(cosh(Y, Ŷ )) (1)
Python 3.6. The experiments were performed on a cluster with
where Ŷ and Y are the prediction and the corresponding ten GeForce GTX 1080Ti GPUs. All the convolutional layers
MAX-DOAS observation of NO2 concentration at the six were initialized randomly, and the biases were initialized
ZHANG et al.: PREDICTION OF VERTICAL PROFILE OF NO2 USING DEEP MULTIMODAL FUSION NETWORK 4101113

with 0. During training, we applied data augmentation by We also used some other evaluation metrics including
adding random Gaussian noise [N (0, 0.005)] to the inputs. Spearman rank correlations, MAE, and RMSE
Besides, all the inputs were normalized by subtracting the n

  
mean pixel value of each channel of the image followed by ρspearman = 1 − 6 di2 / n 3 − n (4)
dividing the corresponding standard deviation. The MF-net i=1
n  
was trained for 100 epochs with batch size of 32, an Adam   1   (h) (h) 
optimizer with β1 of 0.9, β2 of 0.999, and a learning rate M AE h = E Y (h) − Ŷ (h)  = yi − ŷi  (5)
n
of 0.0005. On average, it took 596 s to train the MF-net  2   i=1 2 
with all of the observations in the MAX-DOAS-L data set RM S E = E Y − Ŷ = E E{ Y (h) − Ŷ (h)
on a GeForce GTX 1080Ti GPU. To process one sample,  n 
1  1  (h) 2
6
the MF-net took about 0.22 ms. (h)
= y − ŷi . (6)
6 h=1 n i=1 i

D. Evaluation Metrics
IV. E XPERIMENTAL R ESULTS
To evaluate the performance of our MF-net, we use the In this section, we conducted four experiments to show
Pearson correlation as the evaluation metric. Specifically, the superior performance of MF-net over the WRF-CHEM
the observations of NO2 vertical concentration are noted as (Section IV-A) model. Next, we demonstrated the unique
Y ∈ { yi ; i = 1, 2, . . . , n} and Ŷ ∈ { ŷi ; i = 1, 2, . . . , n} applications of the MF-net (Section IV-B), the ability to
is the corresponding prediction where n is the number of predict NO2 at night, and the more accurate evaluation of the
the observations. The i th observation is represented as yi = transmissions between Beijing and the surrounding cities.
(yi(1) , yi(2) , yi(3) , yi(4) , yi(5) , yi(6) ) for containing six vertical alti-
tudes. Similarly, we use ŷi = ( ŷi(1) , ŷi(2) , ŷi(3) , ŷi(4) , ŷi(5) , ŷi(6) )
for the response predictions. The Ph represents the Pearson A. MF-Net Outperforms the WRF-CHEM Model
correlation between the observations and the predictions of In this section, we first demonstrated that the MF-net out-
NO2 in h th vertical altitudes is defined as performed the WRF-CHEM on the fourfold cross-validation
of the MAX-DOAS-L data set. Then, we draw the same
 
Ph = Corr Y (h) , Ŷ (h) conclusion by comparing WRF-CHEM and the MF-net on
n     the MAX-DOAS-S data set. By testing on the observa-
(h) (h) × ŷ (h) − Ŷ (h)
i=1 yi − Y i tions of 34 CNEMC stations and the satellite data of
=       TROPOMI, we further demonstrated that the MF-net outper-
n (h) (h)
2 n (h) (h)
2
i=1 y i − Y × i=1 ŷ i − Ŷ formed WRF-CHEM on the surface and VCDs prediction of
NO2 , respectively.
h ∈ {1, 2, 3, 4, 5, 6} (2) 1) 3-D Prediction of NO2 on MAX-DOAS-L Data Set:
The fourfold cross-validation experimental results are listed
where Y (h) ∈ {yi(h) ; i = 1, 2, . . . , n} and Ŷ (h) ∈ { ŷi(h) ; i = in Table III and Fig. 3.
1, 2, . . . , n} represent the observations and the predictions of From Table III, we could see that the MF-net outperforms
the NO2 concentration at h vertical altitude, respectively. the WRF-CHEM with smaller MAE, RMSE, and higher
 th
We used Pavg = 16 6h=1 Ph to represent the average Pearson Spearman’s rank correlation on all four sites. As shown in
correlation between the observations and the predictions at the last row of Table III, the average RMSE of MF-net
six altitudes. Since the previous evaluation indicators focused was about 1.29 × 10 molec/cm lower and Spearman’s
11 3

on 2-D prediction, which is not conducive to the comparison rank correlation was 0.176 higher than those of the WRF-
of the model prediction in 3-D, we proposed P3−D as the CHEM. Fig. 3 shows the superior performance of the MF-net
comprehensive indicator to evaluate the model performance over the WRF-CHEM more visually. The darker color bars
in six altitudes. The P3−D was a more appropriate evaluation represented the coefficients evaluated using the WRF-CHEM,
metric for the distribution of concentration along the vertical and the lighter bars were evaluated using the MF-net. It was
direction clear that for any altitude, the light bars were higher than
the dark bars, demonstrating the better performance of the
  MF-net. Furthermore, the Pearson coefficients decreased with
P3D = E{Corr Y, Ŷ }
the increase of the altitude, suggesting that the MF-net perform
6  (h)  
(h)

h=1 yi −Yi × ŷi − Ŷi
better at lower altitude.
1
n
=  2   2 (3) 2) 3-D Prediction of NO2 on MAX-DOAS-S Data Set: In
n i=1 6  (h) 6 (h) this case, the MF-net was trained on all the observations in
h=1 yi −Yi × h=1 ŷi − Ŷi
the MAX-DOAS-L data set and tested on the MAX-DOAS-S
data set, and the results are shown in Table IV and Fig. 4.
where Yi ∈ {yi(h) ; h = 1, 2, . . . , 6} and Ŷi ∈ { ŷi(h) ; h = From Table IV and Fig. 4, we could draw the same
1, 2, . . . , 6} represent the observations and the predictions of conclusion that the MF-net outperformed the WRF-CHEM
the NO2 concentration at a vertical altitude of i th record. with lower MAE and RMSE and higher Spearman’s and
4101113 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

TABLE III
F OURFOLD C ROSS -VALIDATION R ESULTS OF WRF-CHEM AND MF-N ET ON THE MAX-DOAS-L D ATA S ET

smaller and 0.215 higher than those of the WRF-CHEM,

respectively. It was worth mentioning that, since the LC site
was away from the center of the training sites, the performance
of MF-net at the LC site was lower than at the other three
sites.
3) Surface NO2 Prediction as Tested on CNEMC Sites: To
further verify the spatial generalization of MF-net, we tested
MF-net and WRF-CHEM on the observation of 34 CNEMC
stations which are more widely distributed over this region.
As illustrated in Fig. 5, for any one of the 34 stations,
the MF-net achieved higher Pearson correlation than WRF-
CHEM. Majority of the Pearson coefficients evaluated using
the MF-net were more than 0.1 higher than those evaluated
Fig. 3. Comparisons of the Pearson correlation coefficients of the using the WRF-CHEM model, suggesting that the MF-net
WRF-CHEM and MF-net with the MAX-DOAS measurements as tested has good spatial generalization property. From Fig. 5, we can
on the MAX-DOAS-L data set (fourfold cross-validation). The darker bars see that there is a significant improvement at HuaYuan and
represent the performance of WRF-CHEM, while the lighter colors represent
the portions that the MF-net exceeds the WRF-CHEM. YunGang stations and smaller improvements at Yanqing and
Badaling.
4) VCD Predictions Evaluated on the Satellite Data Set: We
compared the performance of the MF-net and WRF-CHEM
on VCD of NO2 within 1km of the ground using the satellite
observations. The results are shown in the Fig. 6.
From Fig. 6, we can see that the MF-net outperforms the
WRF-CHEM at most of the grid point inside of Beijing.
However, we also found in some grids, in the northern
mountain region, that the performance of MF-net was not
good. It was because the stations of MAX-DOAS-L which
we used to train the MF-net were mainly located in the south
and the improvement of MF-net in the north is limited.

B. Applications of MF-Net
Fig. 4. Comparisons of the performance of the WRF-CHEM and MF-net as In this section, we show three applications of the proposed
tested on MAX-DOAS-S data set. MF-net. First, by applying the MF-net, we extended the mon-
itoring time of MAX-DOAS from 15 to 24 h including night.
Second, we evaluated the NO2 transmission at six different
Pearson correlations at most of the altitudes as tested on the altitudes between Beijing and its neighboring cities from Octo-
sites of MAX-DOAS-S data set. And the average RMSE and ber 2016 to November 2019. By comparison, we found that the
Spearman’s coefficients of MF-net are 1.01 × 1011molec/cm 3 WRF-CHEM model overestimated the transmission between
ZHANG et al.: PREDICTION OF VERTICAL PROFILE OF NO2 USING DEEP MULTIMODAL FUSION NETWORK 4101113

TABLE IV
C OMPARISON OF THE P ERFORMANCE OF THE WRF-CHEM AND MF-N ET AS T ESTED ON MAX-DOAS-S D ATA S ET

Fig. 5. Performance of the WRF-CHEM and MF-net as tested on the CNEMC observations during the daytime. The blue bars represent the performance
of MF-net, while the green bars represent the performance of WRF-CHEM. The station numbers are given in Fig. 1.

Fig. 6. Performance of WRF-CHEM and MF-net as tested on the observations of the satellite. (a) and (b) Pearson coefficient for NO2 VCD prediction of
the WRF-CHEM and MF-net, respectively. (c) Difference between the performance of MF-net and WRF-CHEM.

Beijing and Chengde and underestimated the transmission 24 h. To validate the predictions, we evaluated the Pearson
between Beijing and Zhangjiakou. Furthermore, we used the coefficients with the 24-h observations from the CNEMC
MF-net to monitor the distribution and transmission of a NO2 stations [Fig. 7(a)]. The Pearson coefficients were higher than
pollution event in Beijing with a much higher temporal and 0.6 at majority of the CNEMC stations, suggesting the superior
spatial resolution. performance of the MF-net. Fig. 7(a) shows the comparisons
1) MF-Net Extends the Monitoring Time of MAX-DOAS of the WRF-CHEM and MF-net on 24-h observations at
to the Whole Day: With the MF-net, we could predict the 34CNEMC stations. The results demonstrated that the MF-net
vertical profile of NO2 even during night time, allowing outperform WRF-CHEM at all stations in the 24-h evaluation
us to extend the MAX-DOAS monitoring time from 15 to experiments. Fig. 7(b) and (c) shows the visual comparisons
4101113 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Fig. 7. Performance of the WRF-CHEM and MF-net as tested on the CNEMC observations during entire day (24 h). (a) Comparison of the Pearson
coefficient of WRF-CHEM and MF-net evaluated on 34 CNEMC stations during the 24 h. (b) and (c) Visual interpolation results of the time series of MF-net,
WRF-CHEM, and CNEMC observations from April 2, 2018 to April 20, 2018 at Yungang and Yanqing, respectively.

Fig. 8. Comparison of the WRF-CHEM and MF-net based NO2 transmission between Beijing and its neighboring cities from October 2016 to November 2019.
Comparison of WRF-CHEM and MF-net in calculating the monthly transmission of NO2 between Beijing and its neighboring cities.

of the time series of MF-net and WRF-CHEM at two CNEMC As Fig. 8(b) shows the curve between Beijing and Baoding
stations where MF-net got the best and worst performance is always positive, indicating that the transport direction of
by testing on the observations of site, respectively. Compared NO2 is from Baoding to Beijing. The same thing happened
with WRF-CHEM (green), the MF-net (red) is closer to the between Beijing and Zhangjiakou. On the contrary, NO2 was
observations of CNEMC (blue). usually transport from Beijing to Chengde, Tianjing, and
2) More Accurate Transmission Evaluation Than WRF- Langfang. We can also see that Zhangjiakou accounts for the
CHEM: We compared the difference between WRF-CHEM major cause of NO2 pollution in Beijing, and Chengde plays
and MF-net in statistic the transmission of NO2 between a good role in the diffusion export of NO2 in Beijing. One
Beijing and its neighboring cities [Fig. 8(a)] from Octo- phenomenon worth mentioning is that the transmission volume
ber 2016 and November 2019. curve between Beijing and Zhangjiakou is approximately
ZHANG et al.: PREDICTION OF VERTICAL PROFILE OF NO2 USING DEEP MULTIMODAL FUSION NETWORK 4101113

Fig. 9. NO2 VCDs and transmission during the air pollution event in Beijing from November 30, 2017 to December 3, 2017 monitored by MF-net. The
background color indicates the concentration of NO2 VCDs (0 ∼ 1 km). The black arrows represent the wind and the white arrows show the NO2 transmission
along the border of Beijing.

symmetric about the X-axis with the transmission volume December 2, due to the reduced wind speed at 6:00. Finally,
curve between Beijing and Tianjin Langfang. The explanation with the increase of the northwest wind speed, the pollution
is that Zhangjiakou is in the west of Beijing, while the latter slowly disappeared.
are in the south of Beijing. Table V shows the assessment of intercity transport of NO2
In addition, we found that the biggest difference of trans- between Beijing and its neighboring cities in the pollution
mission between MF-net based and WRF-CHEM based was events. It is noted that there are NO2 inflows and outflows
at the border of Beijing and Chengde, as well as the Beijing through the administrative boundaries between Beijing and
and Zhangjiakou. Since we had verified the MF-net out- its adjacent cities. To distinguish them, we added “(+)” and
performed WRF-CHEM in Section IV-A, including testing “(−)” to represent the Beijing’s NO2 incoming and outgoing,
on the satellite VCD observations at the border of Beijing respectively. As shown in Table V, the transmission between
(see Section IV-A4), the MF-net can provide better estimation Beijing and Tianjin calculated by WRF-CHEM is almost
of transmission evaluation. equal as MF-net, similar as between Beijing and Baoding.
But the transmission between Beijing and Chengde calcu-
3) Monitoring the Pollution Events in Continuous Space and
lated by WRF-CHEM was more than MF-net. While for the
Time: Since the MF-net has the capabilities of continuous
transmission between Beijing and Zhangjiakou, the MF-net
monitoring on time and space, we used MF-net to monitor
was higher than WRF-CHEM. To verify whether the MF-net
the distribution and transmission of NO2 in pollution events
exceeds WRF-CHEM in pollution events, we also tested the
with continuous time and space, and the results are shown
two methods on the satellite observations during the four-event
in Fig. 9 and Table V.
periods. The average Pearson of MF-net at all grid points
Fig. 9 shows the distribution and transmission of NO2
inside Beijing is 0.779, higher than 0.743 of WRF-CHEM,
during an air pollution event in Beijing from November 30,
which showed superior performance of MF-net during the
2017 to December 3, 2017. In the beginning, with a northwest
pollution event.
wind, the NO2 in Beijing and its surrounding areas were
thin before the pollution came. From December 1, the wind
direction gradually changed to the southwest, and NO2 was V. D ISCUSSION
transmitted to Beijing from Baoding at 9:00. Then, NO2 started The MF-net applies the meteorology data and the NO2
to accumulate southwest of Beijing from 6:00 to 12:00 on vertical distribution from the WRF-CHEM simulation as the
4101113 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

TABLE V
S TATISTICS OF THE NO2 T RANSMISSION B ETWEEN B EIJING AND I TS N EIGHBORING C ITIES IN
F OUR P OLLUTION E VENTS U SING WRF-CHEM AND MF-N ET

Fig. 10. Comparison of the different altitudes NO2 transmission calculated by MF-net between Beijing and its neighboring cities from October 2016 to
November 2019. (a) Different altitudes transmission of NO2 between Beijing and Chengde. (b) Different altitudes transmission of NO2 between Beijing and
Zhangjiakou.

input, and the sparse but accurately measured NO2 vertical WRF-CHEM, the main reason is that the observations of
distribution as the target. Through the training, the network the assimilation inputs at high altitudes are few than those
establishes a high-dimensional mapping between the simulated at low altitudes. For example, the meteorological measure-
data and the experimental observations. On the other hand, ments are scarce at high altitudes. The emission inventories,
WRF-CHEM is based on a series of physical chemistry laws, the vital information of WRF-CHEM, can only characterize
requiring a sophisticated hyperparameter setting, which lacks the emission at the surface. Since the MF-net applied the
adaptive capabilities to different environments. The MF-net WRF-CHEM outputs as the input, it was not surprising that its
gains the generalization property by training on measure- performance also decreased with an increase in the altitudes.
ments from various sites and allows us to predict the vertical Besides, the measurement error of the MAX-DOAS, used as
NO2 at locations without MAX-DOAS instruments. However, the training label, also increases with height, which further
as shown in Table II, the further the areas are from the four decreases the accuracy of MF-net at high altitudes.
sites used to train the network, the worse the predictions are. We applied both the WRF-CHEM and MF-net to study
One possible explanation is that important factors that affect the NO2 transmission at different altitudes between Beijing
the NO2 vertical distribution could be missing in the MF-net. and its neighboring cities. We found significant differences
For example, the landform or the local traffic pattern may at high altitudes between the two methods. However, since
affect the NO2 vertical distribution and are different for areas there are no direct NO2 vertical measurements along the
further away from the training sites. borders of Beijing current, we could not verify the two
Besides the MF-net spatial generalization property, it also methods directly. Instead, we provided a brief comparison
allows the predictions of vertical NO2 during the night- of WRF-CHEM and MF-net by comparing them with the
time, which lacks MAX-DOAS observations. To validate the satellite’s VCDs observations, which demonstrated the better
predictions, we used the CNEMC 2-D surface NO2 as the performance of the MF-net. We compared the different alti-
references. As shown in Fig. 7(a), the MF-net achieved a tudes transmission of NO2 calculated by MF-net as illustrated
strong correlation with the measurements. Comparing with the in the Appendix (Fig. 10). One interesting find was that the
WRF-CHEM model, the MF-net showed much better perfor- NO2 transmission between Beijing and Chengde is mainly
mance [Fig. 7(c)]. With the MF-net, therefore, we achieved concentrated in the 200 m, while the transmission between
the vertical prediction of NO2 in range of 39.005 − 41.405N Beijing and Zhangjiakou is mainly in the near-surface layer
and 115.005 − 117.905E during the 24 h. and decreased with the altitudes. A reasonable explanation is
The performance of WRF-CHEM and MF-net decreased that the NO2 concentration on the border between Beijing and
with the increase of the altitude (Tables III and IV). For Chengde is low, and the transmission at 200 m is large due to
ZHANG et al.: PREDICTION OF VERTICAL PROFILE OF NO2 USING DEEP MULTIMODAL FUSION NETWORK 4101113

TABLE VI
A BLATION E XPERIMENTS ON MAX-DOAS-L D ATA S ET TO A SSESS THE I MPORTANCE OF MF-N ET I NPUTS . T HE A NNOTATION OF “MF-N ET ( I )”
R EPRESENTED THE MF-N ET WAS O NLY T RAINED W ITH NO2 S IMULATION I NPUT. T HE “MF-N ET (I I )” M ARKED THE N ETWORK WAS T RAINED
W ITH I NPUTTING O NLY THE F IVE M ETEOROLOGICAL VARIABLES W ITHOUT THE NO2 S IMULATION . T HE “MF-N ET ” S HOWED THE
R ESULTS OF T RAINING W ITH B OTH THE M ETEOROLOGICAL VARIABLES AND THE NO2 S IMULATION

large wind speed at high altitude. While the NO2 concentration spatial generalization property of the MF-net in the future.
on the boundary between Beijing and Zhangjiakou was high, The MF-net is not limited to the study of NO2 , it could be
the high concentration of surface NO2 caused larger transmis- used to predict the 3-D vertical profiles of other pollutants,
sion at lower altitudes. such as aerosol and so on.
In Section IV-B3, we proved the better performance of
MF-net during the pollution events using the satellite observa-
VI. C ONCLUSION
tions, and demonstrated a higher transmission between Beijing
and Zhangjiakou, a less transmission between Beijing and The 3-D prediction of NO2 is crucial in understanding the
Chengde. As the previous air pollution research paid more pollution mechanisms, tracking pollutant transport at different
attention to the impact of Baoding on Beijing, we found the altitudes, and formulating the pollution prevention and control
transmission from Zhangjiakou to Beijing dominated the NO2 measures. In this article, we developed an MF-net to achieve
pollution during the pollution events. Our results provided 3-D prediction of NO2 within Beijing and its surrounding areas
scientific data analysis for the pollution control policy making. (115.005 − 117.905E and 39.005 − 41.405N). The MF-net
Currently, we only had four MAX-DOAS observations sites built a map from the WRF-CHEM simulation to MAX-DOAS
with limited data, which prevented us from applying deeper observations and showed superior performance in experiment
neural networks. Furthermore, the stations were within a local results. Compared with WRF-CHEM, the MF-net provided
region and lacked the diversity of different landforms, land a more accurate tool to estimate the 3-D distribution and
uses, and traffic patterns. Therefore, the MF-net did not include transmission of NO2 with high spatial-temporal continuity.
these factors, which limited the application of the network to One interesting find was that previous WRF-CHEM signifi-
locations further away from the training sites. By extending the cantly underestimated the transfer of NO2 from Zhangjiakou
coverage areas of the MAX-DOAS, we expect to improve the to Beijing. The study could help us to better understand the
4101113 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

NO2 pollution in Beijing and provide a scientific basis for [13] A. C. Vandaele, C. Hermans, and S. Fally, “Fourier transform measure-
pollution control policies. ments of SO2 absorption cross sections: II.: Temperature dependence
in the 29 000–44 000 cm−1 (227–345 nm) region,” J. Quant. Spectrosc.
Radiat. Transf., vol. 110, no. 18, pp. 2115–2126, 2009.
A PPENDIX [14] R. Meller and G. K. Moortgat, “Temperature dependence of the absorp-
tion cross sections of formaldehyde between 223 and 323 K in the
To assess the importance of MF-net input including wavelength range 225-375 nm,” J. Geophys. Res., Atmos., vol. 105, no. 6,
meteorological variables and NO2 simulation, we did the pp. 7089–7101, Mar. 2000.
ablation experiments on the MAX-DOAS-L data set with [15] R. Thalman and R. Volkamer, “Temperature dependent absorption cross-
sections of O2 –O2 collision pairs between 340 and 630 nm and at
three different experimental conditions. In the first experiment, atmospherically relevant pressure,” Phys. Chem. Chem. Phys., vol. 15,
the MF-net was trained only with the simulated NO2 as the no. 37, pp. 15371–15381, 2013.
input, without using the meteorological variables. The results [16] X. Tian et al., “Ground-based MAX-DOAS observations of tropospheric
formaldehyde VCDs and comparisons with the CAMS model at a rural
are marked with “MF-net (i)” as shown in Table VI in the site near Beijing during APEC 2014,” Atmos. Chem. Phys., vol. 19,
Appendix. In the second experiment, we trained the MF-net no. 5, pp. 3375–3393, Mar. 2019.
only with the meteorological variables as the input, without [17] R. Shaiganfar et al., “Estimation of the Paris NOx emissions from mobile
MAX-DOAS observations and CHIMERE model simulations during the
using the simulated NO2 . The results were marked with MEGAPOLI campaign using the closed integral method,” Atmos. Chem.
“MF-net (ii).” The third experiment condition was training the Phys., vol. 17, no. 12, p. 7853, 2017.
MF-net with both simulated NO2 and meteorological variables [18] M. Johansson et al., “Quantification of total emission of air pollutants
from Beijing using mobile mini-DOAS,” Atmos. Environ., vol. 42,
marked with “MF-net.” Since the average indicators were a no. 29, pp. 6926–6933, Sep. 2008.
more comprehensive index to evaluate the performance of [19] Q. Hong et al., “Ship-based MAX-DOAS measurements of tropospheric
these methods, we analyzed the results by comparing the NO2 , SO2 , and HCHO distribution along the Yangtze River,” Atmos.
Chem. Phys., vol. 18, no. 8, pp. 5931–5951, 2018.
average indicators. As shown in Table VI, both “MF-net (i)” [20] U. Frieß et al., “MAX-DOAS O4 measurements: A new technique
and “MF-net (ii)” achieved lower average MAE and RMSE, to derive information on atmospheric aerosols: 2. Modeling studies,”
higher Spearman’s and Pearson correlation significantly out- J. Geophys. Res., Atmos., vol. 111, no. 14, pp. 1–20, 2006.
[21] T. Wagner et al., “MAX-DOAS O4 measurements: A new technique
perform the WRF-CHEM. And the “MF-net” achieved the to derive information on atmospheric aerosols—Principles and infor-
lowest average MAE and RMSE and the highest Spearman’s mation content,” J. Geophys. Res., Atmos., vol. 109, no. 22, pp. 1–21,
and Pearson correlation, suggesting the best performance of Nov. 2004.
the training of the MF-net using both simulated NO2 and mete- [22] Z. Javed et al., “Investigating the impact of glyoxal retrieval from MAX-
DOAS observations during haze and non-haze conditions in Beijing,”
orological variables. It was because the two factors containing J. Environ. Sci., vol. 80, pp. 296–305, Jun. 2019.
the complementary information of NO2 . [23] O. Ibrahim, R. Shaiganfar, R. Sinreich, T. Stein, U. Platt, and
T. Wagner, “Car MAX-DOAS measurements around entire cities: Quan-
tification of NOx emissions from the cities of Mannheim and Lud-
R EFERENCES wigshafen (Germany),” Atmos. Meas. Techn., vol. 3, no. 3, pp. 709–721,
Jun. 2010.
[1] H. Liu et al., “A paradox for air pollution controlling in China
revealed by ‘APEC blue’ and ‘parade blue,”’ Sci. Rep., vol. 6, p. 34408, [24] S. Baidar et al., “The CU airborne MAX-DOAS instrument: Vertical
Sep. 2016. profiling of aerosol extinction and trace gases,” Atmos. Meas. Techn.,
vol. 6, no. 3, pp. 719–739, Mar. 2013.
[2] A. J. Zeevaart, “Some effects of fumigating plants for short periods with
NO2 ,” Environ. Pollut., vol. 11, no. 2, pp. 97–108, Sep. 1976. [25] D. Byun et al., “Science algorithms of the EPA models-3 commu-
[3] W. Tan et al., “Estimation of winter time NOx emissions in Hefei, a nity multiscale air quality (CMAQ) modeling system,” Office Res.
typical inland city of China, using mobile MAX-DOAS observations,” Develop., U.S. Environ. Protection Agency, Washington, DC, USA,
Atmos. Environ., vol. 200, pp. 228–242, Mar. 2019. Tech. Rep. EPA/600/R-99/030, 1999.
[4] U. Platt and J. Stutz, “Differential absorption spectroscopy,” in Differen- [26] D. P. Bacon et al., “A dynamically adapting weather and dispersion
tial Optical Absorption Spectroscopy. Berlin, Germany: Springer, 2008, model: The operational multiscale environment model with grid adaptiv-
pp. 135–174. ity (OMEGA),” Monthly Weather Rev., vol. 128, no. 7, pp. 2044–2076,
[5] U. Frieß et al., “Intercomparison of aerosol extinction profiles retrieved 2000.
from MAX-DOAS measurements,” Atmos. Meas. Techn., vol. 9, no. 7, [27] M. Satoh, “Conservative scheme for the compressible nonhydrostatic
pp. 3205–3222, Jul. 2016. models with the horizontally explicit and vertically implicit time inte-
[6] K. Clémer et al., “Multiple wavelength retrieval of tropospheric aerosol gration scheme,” Monthly Weather Rev., vol. 130, no. 5, pp. 1227–1245,
optical properties from MAXDOAS measurements in Beijing,” Atmos. May 2002.
Meas. Techn., vol. 3, no. 4, p. 863, 2010. [28] G. A. Grell et al., “Application of a multiscale, coupled MM5/chemistry
[7] Z. Chen, C. Liu, W. Liu, T. Zhang, and J. Xu, “A synchronous model to the complex terrain of the VOTALP valley campaign,” Atmos.
observation of enhanced aerosol and NO2 over Beijing, China, in Winter Environ., vol. 34, no. 9, pp. 1435–1453, Jan. 2000.
2015,” Sci. Total Environ., vol. 575, pp. 429–436, Jan. 2017. [29] G. A. Grell et al., “Fully coupled ‘online’ chemistry within the
[8] A. Heckel et al., “MAX-DOAS measurements of formaldehyde in the WRF model,” Atmos. Environ., vol. 39, no. 37, pp. 6957–6975,
po-valley,” Atmos. Chem. Phys., vol. 5, no. 4, pp. 909–918, Mar. 2005. 2005.
[9] F. Hendrick et al., “Four years of ground-based MAX-DOAS obser- [30] Y. Zheng, F. Liu, and H.-P. Hsieh, “U-air: When urban air quality
vations of HONO and NO2 in the Beijing area,” Atmos. Chem. Phys., inference meets big data,” in Proc. 19th ACM SIGKDD Int. Conf. Knowl.
vol. 14, no. 2, pp. 765–781, Jan. 2014. Discovery Data Mining, Aug. 2013, pp. 1436–1444.
[10] K. L. Chan et al., “Observations of tropospheric NO2 using ground [31] Y. Zheng et al., “Forecasting fine-grained air quality based on big data,”
based MAX-DOAS and OMI measurements during the Shanghai world in Proc. 21th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining,
expo 2010,” Atmos. Environ., vol. 119, pp. 45–58, Oct. 2015. Aug. 2015, pp. 2267–2276.
[11] K. Chance and R. L. Kurucz, “An improved high-resolution solar refer- [32] Z. Qi, T. Wang, G. Song, W. Hu, X. Li, and Z. Zhang, “Deep
ence spectrum for Earth’s atmosphere measurements in the ultraviolet, air learning: Interpolation, prediction, and feature analysis of fine-
visible, and near infrared,” J. Quant. Spectrosc. Radiat. Transf., vol. 111, grained air quality,” IEEE Trans. Knowl. Data Eng., vol. 30, no. 12,
no. 9, pp. 1289–1295, Jun. 2010. pp. 2285–2297, Dec. 2018.
[12] A. C. Vandaele et al., “Measurements of the NO2 absorption cross- [33] T. Li, H. Shen, Q. Yuan, X. Zhang, and L. Zhang, “Estimating
section from 42 000 cm−1 to 10 000 −1 (238–1000 nm) at 220 K ground-level PM2.5 by fusing satellite and station observations: A geo-
and 294 K,” J. Quant. Spectrosc. Radiat. Transf., vol. 59, nos. 3–5, intelligent deep learning approach,” Geophys. Res. Lett., vol. 44, no. 23,
pp. 171–184, Mar. 1998. pp. 11985–11993, 2017.
ZHANG et al.: PREDICTION OF VERTICAL PROFILE OF NO2 USING DEEP MULTIMODAL FUSION NETWORK 4101113

[34] U. Pak et al., “Deep learning-based PM2.5 prediction considering the Haoran Liu received the Ph.D. degree in environ-
spatiotemporal correlations: A case study of Beijing, China,” Sci. Total mental science from the University of Science and
Environ., vol. 699, Jan. 2020, Art. no. 133561. Technology of China, Hefei, China, in 2019.
[35] S. M. Cabaneros, J. K. Calautit, and B. Hughes, “Spatial estimation of He is a Lecturer at the Institute of Physical Sci-
outdoor NO2 levels in central London using deep neural networks and ence and Information Technology, Anhui University,
a wavelet decomposition technique,” Ecol. Model., vol. 424, May 2020, Hefei. His research interests are regional transport,
Art. no. 109017. and chemical mechanism of air pollution, by ground-
[36] D. Valput, R. Navares, and J. L. Aznarte, “Forecasting hourly NO2 based remote sensing.
concentrations by ensembling neural networks and mesoscale models,”
Neural Comput. Appl., vol. 32, pp. 9331–9342, Aug. 2019.
[37] F. Liu et al., “High-resolution inventory of technologies, activities, and
emissions of coal-fired power plants in China from 1990 to 2010,”
Atmos. Chem. Phys., vol. 15, no. 23, pp. 13299–13317, 2015.
[38] M. Li et al., “Mapping Asian anthropogenic emissions of non-methane Rui Zheng is currently pursuing the Ph.D. degree
volatile organic compounds to multiple chemical mechanisms,” Atmos. with the Department of Precision Machinery and
Chem. Phys., vol. 14, no. 11, pp. 5617–5638, Jun. 2014. Precision Instrumentation, University of Science and
[39] Z. Wang et al., “Elevated dust layers inhibit dissipation of heavy Technology of China (USTC), Hefei, China.
anthropogenic surface air pollution,” Atmos. Chem. Phys., vol. 20, Her major research interests include deep learning,
no. 23, pp. 14917–14932, 2020. intelligent information processing, computer vision,
[40] C. Zhang et al., “First observation of tropospheric nitrogen dioxide and medical image processing and their applications
from the environmental trace gases monitoring instrument onboard the in industry.
GaoFen-5 satellite,” Light, Sci. Appl., vol. 9, no. 1, pp. 1–9, Dec. 2020.
[41] C. Zhang et al., “Satellite UV-Vis spectroscopy: Implications for air
quality trends and their driving forces in China during 2005–2017,”
Light, Sci. Appl., vol. 8, no. 1, pp. 1–12, 2019.
[42] G. Kuhlmann, A. Hartl, H. M. Cheung, Y. F. Lam, and M. O. Wenig, Yizhi Zhu received the bachelor’s degree in envi-
“A novel gridding algorithm to create regional trace gas maps from ronmental science from Anhui Normal University,
satellite observations,” Atmos. Meas. Techn., vol. 7, no. 2, pp. 451–467, Wuhu, China, in 2017. Then, she was recommended
Feb. 2014. to the University of Science and Technology of
China (USTC), Hefei, China, for the Ph.D. degree.
Shulin Zhang is pursuing the Ph.D. degree with the Her research interests are the chemical and phys-
Department of Precision Machinery and Precision ical mechanisms of air pollution by ground-based
Instrumentation, University of Science and Technol- observation and WRF-Chem simulation. Now, she
ogy of China (USTC), Hefei, China. is focusing on the impact of horizontal and vertical
His major research interests include machine transport on air quality, especially regional transport
learning, deep learning, and medical image process- of air pollutants through the higher boundary layer
ing and their application in assisted medical and lower free troposphere.
diagnosis.

Ting Liu received the B.Eng. degree from the Dalian

University of Technology, Dalian, China, in 2018.
Bo Li is pursuing the Ph.D. degree with the School She is pursuing the Ph.D. degree with the School of
of Earth and Space Science, University of Science Earth and Space Science, University of Science and
and Technology of China (USTC), Hefei, China. Technology of China, Hefei, China.
His major research interests include deep learning, Her research interest includes model simulation of
the chemical and physical mechanisms of air pollu- air pollution.
tion, spatial–temporal evolution, regional transport,
chemical mechanism, and the impact of the air
pollution.

Lei Liu received the Ph.D. degree from the Depart- Mingzhai Sun received the Ph.D. degree in physics
ment of Electronic Engineering and Information from the University of Missouri, Columbia, MO,
Science, University of Science and Technology of USA, in 2008.
China (USTC), Hefei, China, in 2019. He is a Professor at the Department of Precision
He is a Post-Doctoral Researcher with the Depart- Machinery and Precision Instrumentation, Univer-
ment of Precision Machinery and Precision Instru- sity of Science and Technology of China (USTC),
mentation, University of Science and Technology of Hefei, China. His research interests include opti-
China (USTC). His research interests include deep cal super-resolution microscopy, biomedical device
learning, statistical learning, evolutionary computa- development, and bioimage informatics, particularly
tion, computer vision, and medical image processing on medical image analysis and deep learning-based
and their applications in industry. algorithm development.

Qihou Hu received the Ph.D. degree in environ-

mental science from the University of Science and Cheng Liu received the Ph.D. degree in environ-
Technology of China, Hefei, China, in 2014. mental physics from Heidelberg University, Heidel-
He is an Associate Professor at the Anhui Institute berg, Germany, in 2010.
of Optics and Fine Mechanics (AIOFM), Hefei He is a Professor at the Department of Precision
Institutes of Physical Science, Chinese Academy of Machinery and Precision Instrumentation, University
Sciences (CASHIPS), Hefei. His research interests of Science and Technology of China (USTC), Hefei,
are the spatiotemporal evolution, regional transport, China. His research interests include the chemi-
and chemical mechanism of air pollution, espe- cal and physical mechanisms of air pollution, the
cially particle and ozone pollution, by ground-based spatial-temporal evolution, regional transport, and
and satellite-based remote sensing observation, and chemical mechanism of air pollution.
numerical simulation.