Papers
Topics
Authors
Recent
Search
2000 character limit reached

CityAQVis: Urban Air Quality Visualization

Updated 11 July 2026
  • CityAQVis is an interactive open-source tool that integrates machine learning models for predicting and visualizing urban air quality using diverse datasets.
  • It utilizes standard scikit-learn algorithms, grid-based data resampling, and hyperparameter tuning to enable comparative analysis of model predictions.
  • Case studies in Bangalore and Delhi illustrate its capacity to reveal spatial pollutant trends and adapt model performance to different urban settings.

Searching arXiv for CityAQVis and closely related air-quality visualization papers to ground the article in current literature. CityAQVis is an interactive machine learning sandbox tool for predicting and visualizing urban air pollutant concentrations at the ground level using multi-source data, including satellite observations, meteorological parameters, population density, elevation, and nighttime lights. It is presented as an open-source system that integrates dataset selection, ML-based spatial modeling and forecasting, and geospatial visualization through a web-based graphical user interface, with a pilot implementation centered on nitrogen dioxide (NO2NO_2) in metropolitan regions (Desai et al., 13 Sep 2025). In related work, the term also appears as a broader shorthand for city-scale air-quality visualization enabled by dense sensing, imagery, and hybrid inference workflows, especially in research on vision-based particulate estimation and urban sensing infrastructures (Chen et al., 2021).

1. Conceptual scope and research setting

CityAQVis addresses a specific gap identified in urban air-quality informatics: interactive tools with seamless integration of forecasting and visualization of spatial distributions of air pollutant concentrations. The software article frames the system as a combined ML and visual analytics environment rather than a stand-alone visualization frontend. Its stated objective is to allow users to build and compare predictive models, visualize model outputs, and derive insights into pollution dynamics at the ground level (Desai et al., 13 Sep 2025).

This positioning places CityAQVis within a broader research trajectory that has treated air-quality visualization as a coupled estimation-and-display problem. HVAQ argues that synchronized dense sensor measurements and wide-view images make it possible to interpolate or estimate pollutant concentrations across the entire field of view and to train image-based algorithms for fine-grained air-quality maps, explicitly identifying this as a key basis for city-scale visualization, or CityAQVis (Chen et al., 2021). ImgSensingNet similarly targets fine-grained air quality monitoring and forecasting through UAV imagery and a ground wireless sensor network, generating 3D, dynamic AQ maps with adaptive sensing (Yang et al., 2019). By contrast, AiR is an Android-based augmented reality application that visualizes various air pollutants present in the locality detected by the user’s GPS or in a locality of the user’s choice, emphasizing in-situ rendering and public awareness rather than model construction or comparative sandbox experimentation (Mathews et al., 2020).

A plausible implication is that CityAQVis should be understood less as a single estimation algorithm than as a unifying software layer for integrating heterogeneous predictors, model comparison, and geospatial rendering in one workflow.

2. Data model, supported inputs, and preprocessing pipeline

The CityAQVis system is described as having three principal stages and a modular architecture supporting data curation, preparation, model training, and interactive visualization. Its supported input data span remote sensing, station observations, meteorology, human activity surrogates, and terrain. These datasets were extracted mainly using Google Earth Engine and spatially resampled via Python’s rasterio (Desai et al., 13 Sep 2025).

Input family Specific data
Satellite observations TROPOMI (Sentinel-5P) Level-3 tropospheric NO2NO_2 vertical column densities
Ground-based measurements Station-based NO2NO_2 concentrations from CCRAQM/CPCB
Meteorological data Rainfall from JAXA GPM GSMaP v6; temperature and wind speed from TerraClimate
Human activity surrogates Population Density from GPWv4/Columbia University; Nighttime Lights from NOAA VIIRS DNB
Terrain Elevation from CGIAR-CSI SRTM 90m DEM

The preprocessing pipeline is explicit. All raster datasets are reprojected or resampled to a uniform 3×33 \times 3 km grid, chosen based on city sizes and data resolution. Monthly or annual means are computed, with the pilot implementation focused on yearly composites for visualization. For ground stations, monthly averages are computed; for gridded predictions, features are extracted at each grid point. Null values are removed before model training, and a standard 70:30 train-test split is used (Desai et al., 13 Sep 2025).

These design choices align with recurrent constraints in urban air-quality datasets. HVAQ emphasizes the importance of synchronized image and sensor data at high temporal and spatial resolution, with measurements collected every second and images captured every 20 minutes, specifically to support evaluation of image-based pollution estimation on both spatial and temporal axes (Chen et al., 2021). AQPDCITY, by contrast, shows that picture-based PM monitoring in big cities remains difficult: nine state-of-the-art methods perform poorly on a 1,500-image urban dataset, highlighting domain shift, scene diversity, and label noise as practical obstacles for urban-scale inference (Zhang et al., 2020). This suggests that CityAQVis’s multi-source fusion strategy is partly a response to the limitations of vision-only approaches in heterogeneous metropolitan scenes.

3. Machine learning integration and analytic workflow

CityAQVis uses ground-level NO2NO_2 concentration (μg/m3\mu g/m^3) as the target variable and treats TROPOMI NO2NO_2, temperature, rainfall, wind speed, population density, elevation, and nighttime lights as input “driving factors.” The supported ML algorithms are Linear Regression, Random Forest Regressor with GridSearchCV for hyperparameter tuning, Support Vector Machine, and Gradient Boosting Regressor with GridSearchCV for hyperparameter tuning. The workflow is interactive: the user selects city, pollutant, year, model type, and feature subset via the GUI; the model is trained on the filtered dataset; and performance metrics are computed as R2R^2 Score, MAE, MSE, MAPE, and RMSE (Desai et al., 13 Sep 2025).

No explicit equations are provided for the CityAQVis estimators because the software wraps standard scikit-learn models. The emphasis is instead on reproducible operational flow: feature extraction, null handling, train-test splitting, and grid-based inference. In that sense, CityAQVis functions as a comparative modeling environment rather than as a proposal for a novel estimator class.

Related systems illustrate alternative modeling choices. HVAQ benchmarks Random Forest Regression, Gradient Boosting Regression, and Support Vector Regression on dense sensor-plus-image data, using dark channel prior transmission for low-altitude images and gray-scale pixel standard deviation for high-altitude images, and reports that GBR achieved the lowest MAE when both images and denser sensors are used (Chen et al., 2021). AQuaMoHo adopts a different problem formulation—AQI label annotation rather than continuous concentration regression—and uses an LSTM with Bahdanau-style temporal attention over thermo-hygrometer readings and spatio-temporal context, with city-specific pre-trained models (Pramanik et al., 2022). ImgSensingNet again differs by casting UAV image inference as AQI scale classification and coupling it to entropy-based semi-supervised learning over a spatial-temporal graph for fine-grained inference and forecasting (Yang et al., 2019).

A plausible implication is that CityAQVis deliberately privileges model interchangeability, parameter transparency, and side-by-side analytic comparison over specialization to a single pollutant-estimation architecture.

4. Visualization logic and interface design

The visualization layer is central to CityAQVis. After training, the selected model infers pollutant concentration for each grid point, creating a gridded prediction map representing a spatial “surface” of predicted values. Because of sparse monitoring data, the system uses grid-based visualization rather than full Kernel Density Estimation, citing computational efficiency and better real-time interaction. A zoom factor limitation on the map makes outputs appear as smooth “blocks,” abstracting the underlying grid granularity (Desai et al., 13 Sep 2025).

The GUI is organized around two main visualization panels that can be configured independently. Each panel contains a map output rendered through Folium or Plotly, model selection and parameters, performance metrics, and validation plots. This design enables direct comparison between two scenarios, such as different years, cities, or models. The user can select the city, prediction year, pollutant, regressors, and included driving factors; choose Folium or Plotly for map rendering; control legend scaling; and inspect actual-versus-predicted validation plots (Desai et al., 13 Sep 2025).

In software terms, the frontend is a Python Streamlit web application. Folium provides tile-based mapping with a fixed legend, whereas Plotly provides a dynamic legend with a user toggle between fixed and global scaling. ML and data processing are implemented through scikit-learn, Pandas, and NumPy (Desai et al., 13 Sep 2025).

This interface design differs substantially from visualization systems that are either purely immersive or purely observational. AiR, for example, renders pollutant molecules in augmented reality, uses the user’s GPS to associate the nearest monitoring station through the Haversine formula, and exposes pollutant-specific informational panels and historical comparisons from Jan 2019 to May 2020 (Mathews et al., 2020). CityAQVis, by contrast, centers on comparative analytical workflows at the map and model level rather than embodied AR rendering.

5. Pilot implementation and empirical behavior

The pilot implementation of CityAQVis uses case studies in Bangalore and Delhi, India. One case study compares predicted NO2NO_2 for Bangalore in 2019, described as pre-COVID, and 2022, described as post-COVID, using Gradient Boosting for 2019 and Random Forest for 2022. The reported finding is that the 2019 model had higher R2R^2, but the 2022 model’s error metrics were lower; spatial maps showed diffusion and a general reduction in NO2NO_20 hotspots from 2019 to 2022, matching known COVID-era emission reductions (Desai et al., 13 Sep 2025).

A second case study compares Bangalore and Delhi for 2019 using the Gradient Boosting model. The reported qualitative pattern is that Bangalore exhibits more localized, moderate NO2NO_21 concentrations and higher model stability, whereas Delhi exhibits higher, more widespread concentrations, with Random Forest performing better in that more complex setting. The paper also highlights specific model validation results: in Bangalore (2019), SVM achieved NO2NO_22 with MAE NO2NO_23; in Delhi (2019), Random Forest performed best with NO2NO_24 and MAE NO2NO_25 (Desai et al., 13 Sep 2025).

The significance of these results lies less in absolute benchmark dominance than in the claim that model performance varied between cities, emphasizing the need for comparative tools. That interpretation is consistent with adjacent literature. AQuaMoHo reports that models are city-specific, that distributions of temperature, humidity, and AQI class boundaries differ by local climate and urban characteristics, and that cross-city transfer learning is not directly addressed (Pramanik et al., 2022). AQPDCITY similarly demonstrates that methods performing on smaller or less diverse datasets may not transfer well to the varied scenes of big cities (Zhang et al., 2020). CityAQVis’s side-by-side design can therefore be read as a software response to heterogeneity in urban pollutant regimes and model behavior.

6. Extensions, deployment, and relation to adjacent CityAQVis paradigms

CityAQVis is described as configurable and extensible. Steps for adding new regions or pollutants include downloading and formatting station and remote sensing data, preparing harmonized input CSV files, and updating web application file paths. The code and case-study data are available at https://github.com/GVCL/CityAQVis, and a live demo is reported at https://cityaqvis-gqj9xf7aqvnj3wkiljvcyf.streamlit.app/. The software is characterized as lightweight, Python- and CPU-based, tested on standard hardware, and deployable on typical laptops (Desai et al., 13 Sep 2025).

The paper further states that although the pilot implementation focuses on annual composites, the workflow allows arbitrary temporal granularity, such as monthly, as input data and processing permit. This suggests an extensible temporal abstraction rather than a fixed annual-only design. It also positions CityAQVis as a general framework adaptable to various pollutants, even though the demonstrated target variable is ground-level NO2NO_26 (Desai et al., 13 Sep 2025).

Related literature shows several directions for such extension. HVAQ provides a public, high-resolution dataset combining point sensor measurements and synchronized images, including NO2NO_27, NO2NO_28, temperature, humidity, GPS coordinates, and timestamps, explicitly to enable high spatial resolution evaluation of image-based air pollution estimation algorithms (Chen et al., 2021). AQuaMoHo proposes fine-grained, localized AQI annotation from a low-cost thermo-hygrometer with GPS plus publicly crawled spatio-temporal information and states that the framework is designed for easy integration into urban visualization and analytics frameworks such as proposed CityAQVis (Pramanik et al., 2022). ImgSensingNet demonstrates an alternative route to city-scale visualization through UAV haze images, selective sensor activation, and entropy-based spatial inference, producing 3D dynamic AQ maps while reducing energy consumption (Yang et al., 2019).

A common misconception is that “CityAQVis” refers exclusively to one visualization modality. The literature instead supports at least two uses. One is the specific software environment introduced in 2025 for multi-source ML-driven urban pollutant prediction and comparative visualization (Desai et al., 13 Sep 2025). The other is a broader research label for city-scale air-quality visualization pipelines enabled by dense sensor networks, wide-view imagery, UAV-ground fusion, or low-cost distributed devices (Chen et al., 2021). Another misconception is that visualization alone is sufficient. Multiple papers indicate that estimation quality depends strongly on sensor density, feature diversity, and urban context, and that visual rendering without robust inference may not generalize well in complex metropolitan settings (Zhang et al., 2020).

In this broader sense, CityAQVis denotes both a concrete software artifact and a convergent research program: the integration of heterogeneous environmental data, predictive ML, and interpretable spatial interfaces for urban air-quality analysis.

Topic to Video (Beta)

No one has generated a video about this topic yet.

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Follow Topic

Get notified by email when new papers are published related to CityAQVis.