These define city static/semi-static data such as demography, morphology, Land Use Land Cover (LULC), urban planning, climatic conditions, industry, and socioeconomic characteristics (Supplementary Table S1). Geographical details of the city are required, as the location of the city determines air pollution dispersion patterns, for example, coastal city or city in a valley. For understanding and modeling the pollutant formation and its dispersion, morphological data such as road type and length, building type, height, and density and trees is required (Cárdenas Rodríguez et al., 2016). Realistic representation of urban LULC also aids in stimulating the planetary boundary layer structure, calculating the thermodynamic and air pollutant transport functions (Sun et al., 2020). Satellite-derived LULC classification, for example, from Landsat and Sentinel, is a viable source for capturing sub-scale granularity and temporal transitions (Saraswat et al., 2017).

A city’s socioeconomic information such as projected population growth, residential area, business hubs, recreation areas, and industry location contribute to the anthropogenic emission source identification (Guttikunda et al., 2013). Geo-referenced ward-wise shapefiles and point shapefiles are part of the city’s geospatial database. Geographical information system (GIS) data such as place of interest, roads, buildings, location layers, and hospital information are important components of the interactive DSS.

These are classified depending on the mode of observation (Supplementary Table S2).

Modeling data is used for three major purposes: understanding the pollution sources, assessing their impacts, and predicting future conditions. These datasets can be long (1–3 months), medium (3–10 days), short (1–3 days), and nowcasting (6–12 h). It includes multiscale (global, regional, and local) weather prediction, real-time and future projection air quality model data, multi-observation atmospheric and chemical data assimilation, health impact, emission factors, and chemical profiles (Baklanov and Zhang, 2020) (Supplementary Table S3). Meteorology data consist of wind speed, direction, turbulence, surface and boundary layer temperature, precipitation, humidity, cloud cover, reflectivity, boundary layer depth, and mixing height affecting pollutant concentration. Boundary layer and lower mixing heights influence the pollutant concentration, and wind speed and direction are responsible for its transport (Oke et al., 2017).

Source apportionment data generated in campaign mode with chemical mass balance models giving source characterization estimates of particulate matter with ions, metals, carbon-specific composition data, and chemical profiles of different emission sources (Pipalatkar et al., 2014; Gargava and Rajagopalan, 2016) form a critical database.

Health exposure assessment model data with mortality, morbidity, and disability-adjusted life years prediction information (Fotopoulou et al., 2016) are the most critical data required for policy decisions and mitigation efficacy analysis. In addition, information on bio air allergens transmissions combined with chemical pollution forms a health database for clinical advisory (Klein et al., 2012).

These databases include static and dynamic data generated by city service departments or national ministries. It consists of fuel type information, vehicle emission standards, sector-wise emission standards, emission factors, daily road segment traffic count, CCTV images, energy demand-supply data, mobility data from smartcards, hospital records, congestion information, different LCS certification data, pollution under control data, smartphone activity data from vendors, Twitter feeds, multi-modal ridership system, route navigation, control measures implementation current and future plans, types of emissions sources to which they are applied and its socioeconomic impact, yearly urban environmental status report, industry regularization reports, air quality Right to Information responses from government departments, and hazardous waste management data (Verma, 2021) (Supplementary Table S4).

IoT with ubiquitous measurements and faster communication enables more granular and better-informed decisions. We consider two types of IoT devices in SmartAirQ: LCS and wearable devices.

LCS provide cheaper and more flexible monitoring options complimenting the expensive reference-grade stations (Hagan et al., 2019). Its flexibility to mount on multiple platforms aids in capturing dynamic pollution characteristics for campaign studies (Apte et al., 2017). SmartAirQ considers LCS for non-regulatory monitoring, hotspots identification, pollution impact assessment on a vulnerable population, awareness, selection of location for regulatory stations, citizen sensing, health exposure assessment, and control measure efficacy analysis (Morawska et al., 2018).

IoT-based wearable devices enable cost-effective epidemiological and individual exposure assessment (Haghi et al., 2018). Wearable devices aid practitioners in the personalized treatment of pollution-related allergies and respiratory diseases (Piedrahita et al., 2014). SmartAirQ will enable the processing of health parameters data derived from crowdsourced data with data privacy and security protocols.

SmartAirQ processes smartphone mobility data to predict traffic-related pollution levels to quantify neighborhood-scale mitigation and geotagged photos for exposure guidance (Donaire-Gonzalez et al., 2016; Gately et al., 2017). They are also effective via media for crowdsourcing LCS data, photos, awareness, and real-time data dissemination through mobile App.

Social media provides a new avenue to gather pollution information (Wang et al., 2017; Charitidis et al., 2019). The best practice example shared in the stakeholder interview, showed that the real time traffic information extracted from the Google mobility data can help city transport planners with better public bus fleet management. The semantic analysis of tweets can be used for mental health impact analysis (Du et al., 2016; Upadhyay and Upadhyay, 2017; Zheng et al., 2019). These datasets, in addition to other data in SmartAirQ DSS, enable city managers to plan, evaluate the mitigation, and build awareness.

Satellite data give concentration information along the column of satellite sensors and on the ground. Geostationary satellite datasets from NASA AQUA/MERRA MODIS and Sentinel-5P, NOAA GOES-R, Hamamury, and CALLIPSO give continuous remote sensed observations of AOD, PM_2.5, PM₁₀, OC, BC, NO₂, SO₂, NH₃, CO, VOC, and dust, aiding in emission trend analysis, filling the observation gaps, sub-scale health exposure, and identifying non-local sources (Wiedinmyer et al., 2011; Kulkarni et al., 2020). NASA Landsat datasets characterize LULC features, source apportionment, and extreme events (Gupta et al., 2006; Jena et al., 2021). SmartAirQ has satellite-based LULC data creation; data fusion of in situ, LCS, drones, ground sun photometer, and satellite data to create granular pollution maps; and data assimilation in models, industry compliance, and health impact assessment (Lau et al., 2019). Recently, 1 km PM2.5 MODIS retrieved datasets for two decades for India will be accessed through SmartAirQ (Dey et al., 2020).

Citizen science brings an opportunity to conduct participative governance, including policy design, awareness, and impact assessment (Mahajan et al., 2020; Lepenies and Zakari, 2021). SmartAirQ enables crowdsourced emission-pollution monitoring through LCS, mobile phone photos, and health impact data acquisition.

With the data mining of scientific and governance data, the urban big data has utility in context-aware data-driven services such as UAQM (Rathore et al., 2018). The UAQM data have big data characteristics of large volume, variety with heterogenous data, velocity with faster data processing, variability due to covariate data dependency, and veracity with data privacy and security (Fazziki et al., 2015; Octaviano et al., 2020). SmartAirQ data architecture builds on the big data principles and technologies such as Hadoop/MongoDB, SPARK, for real-time data processing using cloud computing (Asgari et al., 2017; Huang et al., 2018; Silva et al., 2018).

Hitherto, pollution studies and operational services were limited to in situ observations and coarser models. With the capability to interpret structured and unstructured data, ML methods are used for data mining of large data by training the data and pattern recognition. By building diagnostic, descriptive, prescriptive, and predictive models from it, they enable better actions (Sebestyén et al., 2021). For instance, Liu et al. (2021) developed a correlation analysis of PM10, SO₂, NO₂, and O₃ and showed improvement in prediction data efficiency up to 86% by applying regression models with reference-grade data and other covariate data, such as meteorology using artificial neural network. SmartAirQ uses such ML methods across UAQM lifecycle, namely, random forest in improving LCS calibration (Zimmerman et al., 2018), data fusion (Johansson et al., 2015; Lau et al., 2019), parametrization (Wang et al., 2019), bias correction (Haupt et al., 2021; Xu et al., 2021), extracting information from unstructured data (Zheng et al., 2019), epidemiology (Bellinger et al., 2017), and data sharing within the IoT devices (Ravì et al., 2017).

The computing concept increasingly transitions to “moving analysis to the data” rather than “data to analysis.” SmartAirQ has cloud services with Data as a Service (DaaS), Model as a Service (MaaS), and Platform as a Service (PaaS) connecting to IoT devices, real-time and legacy data, models, and model output to provide last mile services (Liu et al., 2017). It supports data reusability through standardized data workflows, user-specific metadata creation, and sharing. SmartAirQ architecture has service orchestration tools linking with smart city services such as transport, health, economy, and energy. The technology stack has cloud services and processing tools (Supplementary Table S6).

Modeling is an important tool for understanding environmental processes and predicting future conditions. SmartAirQ includes HPC-cloud service with cross-sector, multiscale statistical, and numerical models such as weather, global, regional, and local air quality models and health impact assessment models (Baklanov and Zhang, 2020). The top-down source apportionment method with receptor modeling compares chemical and physical sample properties with emission sources giving the type of the source and its contribution information used in campaign mode (Gargava and Rajagopalan, 2016).

The urban scale dispersion models use downscaled data of initial and boundary conditions, emissions, observational data assimilation, and chemical and physical equation calculations, such as 3D turbulence, convection, boundary layer calculations, transport, and urban heat island influences (Li H. et al., 2020; Ghude et al., 2020; Ramacher et al., 2021). Hyper-local computational fluid dynamics (CFD) models capture pollutant flow around buildings and ventilation coefficient (Santiago et al., 2017). SmartAirQ includes health exposure assessment methods to understand the cause-effect and the disease burden (Smith et al., 2017; Ramos et al., 2018).

Worldwide, DSS with model-based prediction and information systems are gaining importance in air quality governance activities (AirNow; Baklanov et al., 2020; LondonAir; Molina et al., 2019). For example, Zheng et al. (2015) developed an air quality prediction system using machine learning algorithm application on current and past air quality and covariate data such as point of interest, meteorology, traffic flow, and roads. By integrating traditional data sources such as ground observation and chemistry models with satellite, social media, and LCS data, a prototype for hyper-local air pollution DSS (APDSS) for Hong Kong has advanced data analytics and mining methods (Leung et al., 2018). Through highlighting the importance of health risks, the SMURBS ⁶ system has an urban atlas derived from the fusion of satellite data and sub-local exposure assessment data, real-time LCS calibration tool, and emergency action module for industrial accidents with data fusion of CAMS model, citizen observatory, Sentinel-5P and MODIS data, and real-time source apportionment. However, these modules appear as independent components, and an integrated DSS facility is not evident. Furthermore, very few of these go beyond the information portal and include interactive data communication leveraging the 4IR technology for sectorial decisions (Zheng et al., 2015; Leung et al., 2018; Che et al., 2020).

The proposed SmartAirQ framework has not only the best of the above features but also a multi-user secure and interactive science cloud and linked operational service facility harnessing cross-sector pollution mitigation and its integrated assessment, including socioeconomic impacts supported by multiscale modeling and data governance.

The monitoring data quality is improved by calibration methods, such as range test, based on local climatic conditions by mining historical extremes; single sensor and sensor network spatial calibration using linear regression coefficients with reference-grade sensors and land use data (Masiol et al., 2018; Chu et al., 2020); and persistence testing applied by calculating standard deviation below the critical threshold, least square regression (Morawska et al., 2018; Hagan et al., 2019), and multiple regression trees (Simmhan et al., 2019). It also includes CEMS sensor location calibration against standards deployment protocols. This layer has statistical methods to fill gaps in real-time and monitoring data either using historical data mining or co-located station data.

The application query triggers different data processing methods, including pollutant trend analysis, correlation, regression, multivariate data analysis of weather, air quality, transport, health exposure, energy cause-effect analysis, and interpolation to create uniform gridded datasets. The statistical analysis is performed using openAir libraries (Carslaw and Ropkins, 2012). This layer includes control measure efficacy analysis in preventing and reducing polluting incidences, such as “what-if” analysis change in vehicle emission standards and the impact on air pollution levels in a region. Pollution heat maps of all the criteria pollutants are processed for uniform gridding across the city using the inverse-distance-weighing method and calculated from past, current, or future data created from pollution data fusion (Section 5.4).

This plane enables real-time micro-environment exposure calculation based on real-time pollution data (Section 5.4), mobility data, social media, wearable device data, and pushing alerts or advisories through the application service plane.

This layer conducts quality for the LCS as per the crowdsourcing standards (EPA, 2015; Capineri et al., 2016). Social media posts are processed with natural language processing with taxonomy, language filters, and semantic connections using ML methods (Charitidis et al., 2019; Jiang et al., 2019). The classified data is then mapped to create pollution data and impact database. Local climatic zones (LCZ) with GIS are an important consideration due to their ability to detail the localized LULC (natural and anthropogenic) influences such as building height, building cluster, open spaces, vegetation on pollutant concentration, and dispersion (Wu J. et al., 2015; Shi et al., 2019). Standardized LCZ classification methods and tools, such as World Urban Database and Access Portal Tools (WUDAPT), aid in better representation of LULC using crowdsourced morphological data (Ching et al., 2018).

From the observation data, aggregated AQI is calculated for awareness purposes with six categories: good, satisfactory, moderately polluted, poor, very poor, and severe. AQ sub-index and health breakpoints are evolved for eight pollutants (PM10, PM_2.5, NO₂, SO₂, CO, O₃, NH₃, and Pb). AQI are predicted using past and current AQI, past and current weather and model data, social media feeds, and ML (Kosmidis et al., 2018; Castelli et al., 2020). SmartAirQ system has query-based access to sub-local AQI values using GIS visualization and advisory in infographics form for easy understanding of impacted community and citizens, such as the decision regarding school closure during extreme pollution situation in Delhi. ⁷

In response to the application service triggers, this layer transforms the query requirements, connects different data analytics components, allocates the cloud resources, and automates the input-processing-output tasks with service-oriented architecture services (Zanella et al., 2014). It involves multiple job management with scripting and streaming processing for real-time data access, batch processing, and process optimization for DaaS, PaaS, and MaaS. The MaaS connecting multi-sector model workflows with preprocessing, simulations, and post-processing with model data exchanges has OpenMI2.0 framework (Harpham et al., 2019). The workflows run through the public and private cloud (Figure 7).

(1) Public cloud: end users will directly interact with the hosted services that will expose various web applications developed with angular, Django/spring boot framework. It will facilitate users with data visualization using NCL graphics, Vapor, and data analytics using openAir R libraries. OpenStreetMap used in visualization will be exposed via GeoServer, and data can be downloaded via ftp servers.

Various modeling interfaces can be integrated within the UAQM (Kaginalkar et al., 2022). The modeling service layer has global weather forecast models and global, regional, and urban air quality models, such as source apportionment, statistical and simple box dispersion and complex chemical transport models, street-level CFD models, bioweather models, health exposure assessment and socioeconomics impact evaluation models (Baklanov and Zhang, 2020). The models are used for source-concentration estimates, future predictions, and what-if control scenario projections. The model workflow has data containers for data acquisition of 1) global data weather forecast and chemistry model datasets; 2) global, regional, and local EI datasets; 3) air quality observations from regulatory stations, LCS, and satellite 4) LULC data derived from LANDSAT datasets; and 5) weather and air quality observations.

SmartAirQ prototype module consists of regional and urban WRF, WRF-Chem model (Grell et al., 2005), and CFD model OPENFOAM ⁹ . CFD modeling workflows can be invoked in two streams: 1) running online CFD models for the entire city and 2) running CFD models for hotspots as precalculated scenarios and then pattern matching of real-time data using ML methods. It also includes executing the regional-urban-Gaussian dispersion model, such as AERMOD (Cimorelli et al., 2005) to identify multiple hot spots and create sector-specific graded action plans. Users can select further data prediction improvement methods such as hybrid modeling with lognormal/log-logistic statistical distribution model (Gulia et al., 2017) for episodic cases or by running AI models (Section 5.4).

Further, sub-local air quality models can be coupled with health exposure models such as AirQ (AirQ+.) and BENMAP (USEPA, 2014) for estimating exposure and “what-if” scenarios. The model simulated data can be invoked through the query system of the application service plane DSS (Section 5.5).

The model’s accuracy is improved by atmospheric and chemical data assimilation, data fusion methods, and bias correction using hybrid numeric and ML models. The air quality models are executed in ensembles with different initial conditions. The air quality models are supported by evaluation tools for validation with observation data, such as AMET (USEPA, 2016). The compute workloads needed for various simulations and AI/ML engines using several libraries, such as MPI, openMP, TensorFlow, Keras, CUDA, will be executed on HPC clusters (Figure 7).

Efficient data storage is critical in urban environment data management as, every day, huge amounts of data are created and processed. For example, daily three-day forecast Pune city with 500 m WRF model generates ∼17 GB, whereas ten-day forecast for 400 m WRF-Chem model with a larger national capital region of Delhi using chemical data assimilation attributes to 210 GB. This layer is responsible for storing and arranging data with spatiotemporal, sector-specific, impact-specific labeling. It includes a raw data and metadata repository with a periodic storage policy (e.g., few datasets); weather and air quality observations are required as climatology information for past trend analysis or training ML models. For optimal storage, considering the end-user function of the data, lossy or lossless compression techniques are adopted depending on the data accuracy requirements. SmartAirQ leverages distributed data access, wherein data will not always be stored at a centralized location but organized on multiple physical or virtual infrastructures. The repository also includes information on apps, different model configurations, scripts, and data connection between different models as workflows. Data management services will have a big data framework using Hadoop/MongoDB, which will house the data needed for SmartAirQ. It will also have extraction transformation load (ETL) channels implemented in spark, python, to ingest data into the data store (Figure 7).

ML-based calibration methods such as support vector machine, random forest, K-nearest neighbor, and geospatial kriging are applied to LCS data and covariate data to improve the accuracy (Maag et al., 2018; Zoest et al., 2020).

Overcoming the limitations of various monitoring methods, data fusion processes monitoring data from in situ data, campaign study data, satellite data, chemical transport models, and smart city air quality LCS data using ML methods to create gridded datasets of PM_2.5, PM₁₀, NO₂, CO₂, SO₂, and O₃ (Li et al., 2017; Schneider et al., 2020; Viqueira et al., 2020).

Data fusion is executed in two ways: offline and real-time. Generally, in the offline mode, data used in specific campaigns are fused to create a newer dataset. For example, LCS, drones data, and mobile LCS data on vehicles are combined to create an emission source spatiotemporal map. These datasets are then used to train the dataset for real-time feature extraction in data fusion methods (Lau et al., 2019; Geng et al., 2021). Then, the data created by data fusion method is compared with the in-situ station data.

Pollution reduction is dependent on emission source-concentration information. Building scale LULC data is prepared using satellite data and ML methods (Kerins et al., 2020). This layer performs spatial (roads, buildings, place of interest, and types of roads) and temporal classifiers based on artificial neural network and regression methods on cross-variate data training models to improve pollutant concentration predictability (Zheng et al., 2015). In this layer, ML methods K-nearest neighbor and long short-term memory are applied for traffic density assessment by vehicle count from traffic sensor data, trajectory analysis, and visual classification from CCTV images from smart city command and control data services for congestion prediction (Majumdar et al., 2021). These datasets are linked to pollutant level datasets and GIS data for less polluted route navigation and emission zone surcharge advisory on the mobile app.

LCS and social media data are used to identify emission sources and control measures. This layer includes content extraction and pattern recognition using natural language processing and semantic analysis using ML methods (Zheng et al., 2019). Data are then correlated to pollution concentration data to predict future AQI and health impact assessment data to be used by city managers and by health practitioners (Yan et al., 2019). Satellite-derived emission source data and pollutant and LULC data are correlated with CEMS data for compliance monitoring (Kurinji and Ganguly, 2020).

This plane has provision for health impact data processing using hospital records, traffic sensor data, mobile phone activity data, wearable device data, ridership data, real-time and historical meteorology data, social media, and gridded pollution data generated from the data fusion layer and using ML methods such as random forest, support vector machine, and multiple regression methods giving pollution exposure information at a location and time (Hu et al., 2014; Masiol et al., 2018).

ML methods are applied to air quality models, such as WRF-Chem, to improve predictability by bias correction and for faster processing (Ma et al., 2020). The processing is executed using Keras, TensorFlow, PyTorch, NumPy, python scripts, and CUDA environment.

The data processing pipelines involve various tools and libraries, as shown in Figure 7 (Supplementary Table S6).

Data visualization with GIS is an inherent component of the city’s digital platform and DSS. Its usage runs across all the components of SmartAirQ, from EI development to dissemination services (Dalvi et al., 2006; Badach et al., 2020). WebGIS allows full view of the city, standardized data access, interactive location-specific query, and navigation (Gkatzoflias et al., 2013). SmartAirQ architecture considers it a vertical plane running across all other service planes.

Visual representation of data is triggered by user query or sector-specific action in the form of infographics, plots, 3D animation, augmented reality and virtual reality (AR/VR), time-series graphs, trend analysis, correlation, exceedance, prediction information, industry compliance reports, and AQI display (Elbir et al., 2010; Li W. et al., 2020). The dynamic visualization includes interactive navigation displaying pollutant information along all the routes with an expected time of arrival information. The application service has a restful API for data exchanges. The scientific plots are viewed through Vapor/NCL.

For pollution mitigation and its efficacy assessment, large data must be brought together, linked, and inferred in the context in which it is applied, which requires an efficient database management system maneuvering cross-sector data and metadata.

In this use case, the control measures selection process is instrumented by a local environment manager, considering the business as usual scenario or future scenario projection data analytics. It accesses the pre-simulated future projections using dispersion models.

Emission factors and EI activity data generated from source apportionment studies are uploaded by researchers through the data acquisition plane (Figure 8). The system applies data wrangling and data classification methods by data mining of source categories and sub-categories. The city environment manager selects business as usual scenarios and control scenarios. The data are processed in the data processing plane with a calculation of emission load, reduction factor, and applicability ratio. Based on the query and user-selected pollutant type, source type from the area, point or line, scenario type, such as the closure of construction work in a fixed time or banning diesel vehicles, emission factor selection, and analytics mode, such as emission load in kg/year or ton/day, a cumulated scenario is generated (Figure 9). The backend calculations are executed using WRF-Chem and AERMOD models projected simulations using cloud resources, and analytics is conducted using R libraries.

The total % reduction is displayed either as plots or GIS-based heat maps (Figure 9). In the decision chain, the manager links the data with energy, transport, traffic, infrastructure department, and state pollution control board members. The concerned stakeholders view the applicability and feasibility of control measures with interactive pollutant, source, control scenarios selection, comparison between two locations, and so forth. The selected control measures are then implemented by respective stakeholders.