21 Engineering and Technology Journal for Research and Innovation (ETJRI)

ISSN 2581-8678, Volume III, Issue II, July 2021

Application of Geospatial Artificial Intelligence in

mapping of air pollutants in urban cities
Sushree Sangita Mishra1, Deepa Parasar2*
1
Associate Professor, CE, ASET, Amity University Maharashtra, [email protected]
2*
Associate Professor, CSE, ASET, Amity University Maharashtra, [email protected]

Abstract— People from most of the countries particular, has a significant effect on
across the globe have been distressed due to air human health due to the exposure to toxins
pollution induced acute health problems, which
and particulates, which have piqued the
emphasize on the need of air pollution
monitoring. Remote sensing of high resolution scientific community's interest in air
satellite imageries is a useful tool that facilitates pollution and its consequences. The
a better interpretation and analysis of the combustion of fossil fuels, deforestation,
concentration of air pollutants. Satellite waste from factories and industries,
imageries are helpful to quantify and map air
domestic heating, and natural disasters are
pollution because they are able to provide
synoptic view of large areas. Different satellite the major sources of air pollution. Air
sensors with various ranges of temporal, spatial pollution is one of the world's most critical
and radiometric resolution offer the opportunity issues. The consistency of the air varies
to estimate the concentration of harmful air dramatically over time and space. Because
pollutants such as carbon monoxide, nitrogen
of the strong temporal resolution of
dioxide, ammonia, sulphur dioxide, volatile
organic compounds, and surface particulate different satellite sensors, satellite data can
matter (PM2.5). In addition to mapping of the be used to quantify and map air pollution
air pollutants to check its detrimental effects on because it can provide full views of vast
human health, it is also important to predict the areas with just one image on a regular
future concentrations by developing models
basis. Different air pollutants, such as
using different simulation methods. These
simulated models may help the government nitrogen dioxide (NO2), sulphur dioxide
officials and authorities for policy (SO2), ammonia (NH3), carbon monoxide
implementation towards the control of air (CO), certain volatile organic compounds
pollution levels. Though many research works (VOCs), and aerosol optical depth (AOD),
are available on different approaches to model
from which surface particulate matter is
air pollutant levels, recent advancements in deep
learning techniques are more promising with extracted, can be estimated using a range
accurate forecasting results. This paper presents of satellites. To construct a surface grid or
a review of the remote sensing technique in contour chart, spatial interpolation
mapping and prioritizing the concentration of techniques are used. Using known
air pollutants. In addition to this, a description
concentrations at fewer points,
of the deep learning architectures to develop
models on air pollution data obtained from the interpolation processes estimate
satellite sensors is also discussed. concentrations in the study region. A
geographic information system (GIS) is a
Keywords: Air pollution, Remote sensing, Deep software tool for storing, analysing,
learning, Sensors, Spatial interpolation.
managing, and presenting all forms of
I. INTRODUCTION spatial or geographical data, which
produces solutions to spatial problems
Environmental pollution issues such as when combined with the expert opinion of
water, noise, and air pollution are the GIS user or analyst. Various
emerging as cities grow economically and researchers have used GIS techniques to
technologically. Air pollution, in investigate the spatial and temporal
22 Engineering and Technology Journal for Research and Innovation (ETJRI)
ISSN 2581-8678, Volume III, Issue II, July 2021

distribution of contaminants. One of the stations to track air quality, but the
aims of this project was to measure the information gathered by these stations is
temporal and spatial variation of air inadequate to map air pollution because air
pollution across the country. An air quality quality is both spatially and temporally
monitoring network system is required to variable. Since satellite images have a high
collect data on SO2, NO2, and SPM temporal resolution, they can cover a wide
emissions concentrations. The point data area in a single image, making satellite
files were also developed for the Mumbai remote sensing a powerful tool for air
region's air quality monitoring stations. At pollution control [23].
seasonal and annual time scales, air quality Several studies have used satellite
concentration data for SO2, NO2, and images to track and chart air contaminants.
SPM was assigned to the appropriate point For example, OMI was used to monitor
form file. At weekly, seasonal, and annual NOx emissions in Colombia [7], the mid-
temporal scales, the collected air quality Atlantic United States [5], the United
data set can be fed as an attribute of points States [11] [8] [13] and Europe [16].OMI
form files. The current study's approach was also used to measure SO2 emissions
aims to discuss and review effective in North America [9]. Abdel Satter [4]
models that can be used to forecast current looked at the capabilities of various
Particulate matter concentration levels. satellite remote sensing technologies for
Such forecasts result in an estimation of air calculating air contaminants, as well as the
quality (in terms of Particulate Matter methods for processing and accessing
levels) that can be useful to local decision satellite data for pollutant concentration
makers in charge of the day-to-day mapping. According to the researcher,
operational implementation of relevant different spectral resolutions of space
legal structures, such as issuing pollution instruments allow for the detection of
exceedance alerts and enforcing effective various species of air pollutants. Air
steps. Machine learning (ML) and Deep pollution measurements from space are
Learning models can be used to analyse useful for monitoring air quality and
the data obtained by RS and to meet the studying the long-term pattern of
need for more accurate forecasting in atmospheric pollutant concentrations.
various scientific areas and domains. ML Support vector regression (SVR) [3] is
algorithms, in particular, have been used to predict pollutant and particulate
commonly used to predict air quality. levels and calculate the air quality index
(AQI). Out of all the alternatives tested,
II. LITERATURE SURVEY the radial basis function (RBF) was the
Owing to prolonged exposure to air type of kernel that allowed SVR to make
contaminants such as nitrogen dioxide the most accurate predictions. The results
(NO2), particulate matter (PM), ozone showed that SVR with the RBF kernel
(O3), and sulphur dioxide, millions of could reliably predict hourly pollutant
people across the world are suffering from concentrations such as carbon monoxide,
severe health problems (SO2). According sulphur dioxide, nitrogen dioxide, ground-
to the World Health Organization [10], air level ozone, and particulate matter 2.5, as
pollution killed over 4.2 million people in well as the hourly AQI for California.
2018.Furthermore, according to WHO, Research was conducted [22] to create
about 90% of the world's population lives regression models that can predict the
in areas where air pollution levels exceed spatial and temporal variability of PM10 in
WHO air standards [10]. As a result, cyprus (particulate matter with an
global attention was drawn to air pollution. aerodynamic diameter less than 10 m)
Many countries have set up ground ground measurements, Aerosol Optical
23 Engineering and Technology Journal for Research and Innovation (ETJRI)
ISSN 2581-8678, Volume III, Issue II, July 2021

Depth retrievals from satellite, and the layer to consider the time dependence of
prevailing synoptic conditions developed pollutants in forecasting PM2.5
by Artificial Neural Networks. An concentration. IoT [19] has been
adequately developed neural classification incorporated for measuring a variety of
system clustered synoptic charts, Aerosol city parameters at Antwerp, Belgium,
Optical Depth data from the Aqua satellite, including air pollution. To deal with
and ground measurements of particulate temporal and spatial resolution, the
matter at the heart of the forecasting problem is addressed as matrix completion
system. Statistical models for forecasting on graphs problem and relies on variation
emission levels were created using the graph auto encoders to propose a deep
above tools. A similar study was learning solution for the estimation of the
conducted [24] in Equador to model unknown air pollution values. A new deep
PM2.5 urban emissions using machine learning method[17] was proposed based
learning and meteorological parameters. for air quality forecasting model, namely
The classification model's findings DFS (Deep Flexible Sequential) a hybrid
indicate that low (10 g/m3) versus high of LSTM and CNN to forecast PM10 as
(>25 g/m3) and low (10 g/m3) versus the target pollutant through the work in the
moderate (10–25 g/m3) PM2.5 region of Istanbul Turkey between 2014
concentrations are reliably classified. A and 2018. A Deep Multi-output LSTM
regression analysis indicates a better (DM-LSTM)[25] neural network model
estimate of PM2.5 when climatic has been proposed by incorporating
conditions become more extreme (strong features of three deep learning algorithms
winds or high levels of precipitation). (i.e., mini-batch gradient descent, dropout
When climatic conditions become more neuron and L2 regularization) to configure
severe, a regression analysis suggests a the model for extracting the key factors of
better estimate of PM2.5 (strong winds or complex spatio-temporal relations as well
high levels of precipitation). as reducing error accumulation and
Across the globe it has been identified propagation in multi-step-ahead air quality
that adoption of AI and machine learning forecasting at Taipei, City of Taiwan. This
in dealing with hazards of air pollution is is advancement over the Shallow Multi-
on its frontiers. Deep learning has been output Long Short-Term Memory (SM-
described as a promising field in the LSTM) model which is suitable for
situation, with the potential to increase the regional multi-step-ahead air quality
resolution and precision of predictions in forecasting, but encounters spatio-temporal
data-poor scenarios, according to the instabilities and time-lag effects.
literature. It has been argued that low- and
middle-income countries with minimal III. METHODOLOGY
monitoring and surveillance infrastructure Geospatial artificial intelligence has
may depend on image-based convolutional been successfully used in many disaster
neural networks [21] as a cost-effective related problems and agricultural studies
method of estimating local environmental by many researchers. It has also been used
exposures. Choosing Shanghai, China as for air quality mapping and modelling of
the target city [2] combined two kinds of air pollutants in urban cities. This method
deep networks for extracting features in encompasses two techniques viz.
both spatial and temporal dimensions, Geospatial Technology and artificial
Convolutional neural network is taken as intelligence. The geospatial technology
the base layer, automatically extracting can be used for spatio- temporal mapping
features of input data. A long short-term of air pollutants and prioritization of
memory network is used for the output locations based on this analysis. The
24 Engineering and Technology Journal for Research and Innovation (ETJRI)
ISSN 2581-8678,
8678, Volume III, Issue II, July 2021

Artificial intelligence
ligence techniques viz. Deep the overallll surface curvature [20].
[2 This
Learning techniques, ANN, CNN etc. are enables a smooth surface formed exactly
then applied to forecast the concentration accommodating all the points. This
of the pollutants to install air purifiers in interpolation method uses a different form
the prioritized locations. A generalized of polynomial interpolation and is best for
methodology flowchart chart is mentioned surfaces that change gradually. It is the
below (Fig. 1). best tool for representing gradually
A. Mapping of Air Pollutants varying surfaces with small errors because
it can forecast valleys and ridges in the
The concentration of air pollutants dataset.
specifically SO2, NO2, and SPM are
obtained from the either from the
government organizations or can be
extracted from satellite images using
image processing software. The
concentration of air pollutants are mapped
and spatially interpolated using different
GIS based spatial interpolation techniques
viz. Kriging method, splines, Inverse
distance weighted (IDW) method etc.
These techniques have been used
successfully in air pollution studies [1[12],
[15]. IDW method is used when hen density of
the set of points is high so that a local
surface variation can be obtained for the
analysis. It uses a linear weighted
combination of data to compute the grid
values of any random parameter. It
depends on the distance between the
sampling sites and the location where the
interpolation is needed [14].
]. Points that are
closer to the interpolation point for SPM
would have more weight. The IDW
method is an accurate interpolation
technique that executes the condition that
the forecasted value of a point is more Fig. 1. Methodology Flowchart
influenced by nearer known points than by
those which are far away.. The general Kriging is a spatial interpolation
equation for the IDW method is: technique that uses a Gaussian method
with prior covariance to model the
∑
interpolated values [1 1]. This stochastic
z = (1) approach can be used in a variety of
∑ situations, including disaster prevention,
Where z is the computed value at point geochemistry, and pollution modelling
0, 𝑧 is the z value at known point I, 𝑑 is [6].It's
].It's based on the idea that the distance
the distance between point I and point 0, s between different sample points poi is
is the number of known points used in spatially associated, indicating surface
interpolation,, and k is the specified power. variation. Random noise is reduced using
Spline uses a mathematical function to this form. It calculates the values for all
compute unknown values by minimizing points within a predetermined range. It
25 Engineering and Technology Journal for Research and Innovation (ETJRI)
ISSN 2581-8678, Volume III, Issue II, July 2021

forecasts values from observed data using After pre-processing the images can be
a more sophisticated weighted average used to simulate a model to forecast the
approach. Mathematically, this function is concentration of air using Artificial
represented as Intelligence and machine learning
techniques. As a sub field of Machine
learning, Deep learning has gained
---- (2) tremendous impetus in finding
applications in various fields like Natural
Language Processing (NLP), healthcare,
where u and ui are location vectors for bioinformatics and exposure science over
computation of point and one of the these years. Its ability of employing
neighbouring data points, indexed by i; n numerous distinct deep neural network
(u) is the total number of data points in the architectures has made it a popular choice
local neighbourhood used for estimation of to get state of the art results in varied
Z*(u); m(u) and m(ui) are expected values applications. With a mathematical
(means) of Z(u) and Z(ui); and λi(u) is the foundation of algebra and calculus it
Kriging weight assigned for datum Z(ui) implements neural networks with more
for estimation of location u; the same than a single hidden layer of neuron. Deep
datum will receive a different weight for learning model building will be relying on
different estimation location. these 5 important steps as mentioned
below:
B. Air Pollutant Forecasting
Defining the architecture: For defining
Multidate images are needed for the architecture the nature of problem is to
forecasting of the air pollutants. All be studied. Due to the nature of the
images are pre-processed using the problem studied, Convolution Neural
following steps. Networks (CNN) happens to be a popular
choice to carry out work in image
 Acquisition of Images – It involves segmentation and image classification on
acquiring and storing of images to which detailed prediction work can be
database. Further converting it into carried out [18]. The model building
a variable, creation of load folders structure used for the entire deep learning
containing images into arrays. architecture is Sequential Models,
 Resizing Images – In order to Functional API, or custom architecture
bridge the mismatch in the which can be defined.
requirement of our model and the Compilation of the model – To
actual image captured, image configure the model for the fitting/training
resizing is necessary. process compilation of the model should
 Denoising the images – Unwanted be done. Few of the critical components of
image noises are to be curbed. the training procedure get defined for the
Smoothening of images is a evaluation procedure during the
requirement in order to enhance compilation step. Due to the nature of the
different scales image structures. problem we will be subjected to losses
 Segmentation and Morphological which are to be defined at this step, we
changes – Image segmentation may also need to decide on the optimizers and
be required for separating the metrics such as accuracy and classification
foreground from the background related metrics are to be defined.
for extracting useful features. Edge Step 3: Fitting the model – To fit the
Smoothening is required for model on the training dataset happens to
implementing morphological be an important procedure. It trains the
changes.
26 Engineering and Technology Journal for Research and Innovation (ETJRI)
ISSN 2581-8678, Volume III, Issue II, July 2021

model for a fixed number of epochs department and other Government and
(iterations on the dataset).During the entire private agencies to forecast air pollution
training procedure the fitting step must be levels so that further action can be taken
continuously evaluated. It is essential to accordingly.
make sure the model being trained is
performing well with improving accuracies REFERENCES
as well as a reduction in the overall loss. [1]. A. Satter, “Monitoring Air Pollution Using
Over-fitting the model will also be Satellite Data,” in Proc. International
Conference on Industrial Engineering and
avoided. Tools like Tensor board will be Operations Management Riyadh, Saudi
handy for such a procedure. Arabia, 2019.
Step 4: Evaluating and Making [2]. T.C. Bailey, and A.C. Gatrell. Interactive
Predictions — evaluating the performance Spatial Data Analysis. London, UK:
of the deep learning model in real Longman. 1995.
[3]. S. Beirle, K. F. Boersma, U. Platt, M. G.
scenarios of our application is important Lawrence, and T. Wagner, “Megacity
and will be carried out. We need to ensure Emissions and Lifetimes of Nitrogen
that the predictions made by our model on Oxides Probed from Space,” Science, vol.
the test data (split at the beginning of the 333, pp. 1737-1739, 2011.
pre-processing step) are taken into [4]. K. F. Boersma, D. J. Jacob, E. J. Bucsela,
A. E. Perring, R. Dirksen, R. M. Yantosca,
consideration for the purposes of R. J. Park, M. O. Wenig, T. H. Bertram,
validating the effectiveness of the trained and R. C. Cohen, “Validation of OMI
model. Further Random tests will further tropospheric NO2 observations during
substantiate the effectiveness on untrained INTEX-B and application to constrain
data. NOx emissions over the eastern United
States and Mexico,” Atmos. Environ., vol.
Step 5 Deploy the model – As a final 42, no. 19, pp. 4480-4497, 2008.
stage of any model construction it has to [5]. M. Castelli, F. M. Clemente, A. Popovic,
be deployed. We plan to deploy it as an Silva, S., Vanneschi, L, A Machine
android and web application building it Learning Approach to Predict Air Quality
with either flask, Django, or any other in California, Hindawi, Vol. 2020, Article
ID
similar framework. 8049504,https://doi.org/10.1155/2020/804
9504, 2020.
IV. CONCLUSION [6]. R. L. Curier, R. Kranenburg, A. J. S.
The rapidly increasing concentration of Segers, R. M. A. Timmermans, and M.
Schaap, “Synergistic use of OMI NO2
air pollutant levels in urban cities is tropospheric columns and LOTOS–
alarming; hence air pollution monitoring EUROS to evaluate the NOx emission
study is imperative. Satellite remote trends across Europe,” Remote Sensing of
sensing technique enables the Environment, vol.149, pp. 58-69, 2014.
quantification of the concentration of air [7]. J.K. Deters, R. Zalakeviciute, M.
Gonzalez, Y. Rybarczyk, “Modeling
pollutants and locations can be prioritized PM2.5 Urban Pollution Using Machine
based on GIS based spatial interpolation Learning and Selected Meteorological
techniques. Moreover, having considered Parameters,” Journal of Electrical and
the calamitous effects of air pollutants on Computer Engineering, Article ID
human health, it is important to forecast air 5106045,https://doi.org/10.1155/2017/510
6045, 2017.
pollutant concentrations by simulation of [8]. T. H. Do, E. Tsiligianni, X. Qin, J.
models. This study was aimed at providing Hofman, V. P. La Manna, W. Philips, and
a systematic review of the satellite remote N. Deligiannis, “Graph-Deep-Learning-
sensing techniques in mapping and Based Inference of Fine-Grained Air
prioritizing air pollutant levels and also to Quality from Mobile IoT Sensors.” IEEE
Internet of Things Journal, vol.7, no. 9,
predict the concentration of air pollutants. 8943-8955. (2020).
It will help people in the meteorological
27 Engineering and Technology Journal for Research and Innovation (ETJRI)
ISSN 2581-8678, Volume III, Issue II, July 2021

[9]. V.E. Fioletov, C.A. McLinden, N. Satellite, Ground-Based, and Synoptic

Krotkov, M.D. Moran, K. Yang, Data, Elaborated with Regression
“Estimation of SO2 emissions using OMI Models,” Hindawi, vol. 2017,
retrievals,” Geophys. Res. Lett. Vol. 38, https://doi.org/10.1155/2017/2954010.
L21811, 2011. [20]. D. Qin, J. Yu, G. Zou, R. Yong, Q.
[10]. D. L. Goldberg, L. N. Lamsal, C. P. Zhao, and B. Zhang, “A novel combined
Loughner, W. H. Swartz, Z. Lu, and D. G. prediction scheme based on CNN and
Streets, “A high-resolution and LSTM for urban PM 2.5 concentration.”
observationally constrained OMI NO2 IEEE Access, vol. 7, pp. 20050-20059,
satellite retrieval,” Atmospheric Chemistry 2019.
and Physics, vol. 17, pp. 11403–11421, [21]. S. Weichenthal, M. Hatzopoulou, and M.
2017. Brauer, “A picture tells a thousand
[11]. J. F. Grajales, and A. Baquero-bernal, exposures: opportunities and challenges of
“Inference of surface concentrations of deep learning image analyses in exposure
nitrogen dioxide (NO2) in Colombia from science and environmental epidemiology”
tropospheric columns of the ozone Environment international, vol. 122, pp. 3-
measurement instrument (OMI),” 10, 2019.
Atmósfera, vol. 27, no. 2, pp. 193–214, [22]. D.W. Wong, L. Yuan, and S.A. Perlin,
2014. “Comparison of spatial interpolation
[12]. H. Greiner, “A survey on univariate data methods for the estimation of air quality
interpolation and approximation by splines data.” J. Expos. Anal. Environ. Epidemiol.
of given shape.” Math. Comput. Model. vol.14, no. 5, pp. 404–415. 2004.
Vol.15, no. 10: pp. 97– [23]. World Health Organization (WHO). Air
108.doi:10.1016/0895-7177 (91)90094-N. Quality and Health. Fact Sheet N 313.
1991. Updated May 2018.
[13]. D.A. Griffith, Advanced Spatial [24]. L. Zhang, D. J. Jacob, K. F. Boersma, D.
Statistics. Dordrecht, The Netherlands: A. Jaffe, J. R. Olson, K. W. Bowman, J. R.
Kluwer Academic, 1988. Worden, A. M. Thompson, M. A. Avery,
[14]. J.L. Gunnink, and P.A. Burrough R. C. Cohen, J. E. Dibb, F. M. Flock, H.
“Interactive spatial analysis of soil E. Fuelberg, L. G. Huey, W. McMillan, H.
attribute patterns using exploratory data B. Singh, and A. J. Weinheimer,
analysis (EDA) and GIS.” In Spatial “Transpacific transport of ozone pollution
Analytical Perspectives on GIS, ed. I. and the effect of recent Asian emission
Masser and F. Salge,. New York, N Y: increases on air quality in North America:
Taylor & Francis. pp. 87–99. 1996. an integrated analysis using satellite,
[15]. D. G. Hadjimitsis, K. Themistocleous, aircraft, ozonesonde, and surface
and A. Nisantzi, “Air Pollution observations,” Atmos. Chem. Phys., vol.
Monitoring Using Earth Observation & 8, pp. 6117-6136, 2008.
GIS,” Air Pollution. Rijeka: IntechOpen, [25]. Y. Zhou, F. J. Chang, L. C. Chang, I. F.
2012. Kao, and Y. S. Wang, “Explore a deep
[16]. D.K. Jha, M. Sabesan, A. Das, N.V. learning multi-output neural network for
Vinithkumar, and R. Kirubagaran, regional multi-step-ahead air quality
“Evaluation of interpolation technique for forecasts”, Journal of cleaner production,
air quality parameters in Port Blair, India.” vol.209, pp. 134-145, 2019.
Univ. J. Environ. Res. Technol. vol. 1, pp.
301–310. 2011.
[17]. K. Kaya, and Ş. G. Öğüdücü, “Deep
Flexible Sequential (DFS) Model for Air
Pollution Forecasting.” Scientific Reports,
vol. 10, no. 1, pp. 1-12, 2020.
[18]. X. Li, L. Peng, Y. Hu, “Deep learning
architecture for air quality predictions.”
Environ Sci Pollut Res vol. 23, pp. 22408–
22417, 2016
https://doi.org/10.1007/s11356-016-7812-
9.
[19]. S. Michaelides, D. Paronis, A. Retails, F.
Tymvios, “Monitoring and Forecasting
Air Pollution Levels by Exploiting