A Visual Analytics Approach for Station-Based Air Quality Data

Yi Du, CuiXia Ma, Chao Wu, Xiaowei Xu, Yike Guo, Yuanchun Zhou, Jianhui Li

TL;DR

This paper introduces AirVis, a visual analytics system designed to tackle the complexity of multi-dimensional, spatio-temporal air quality data. By integrating map-based views, calendar views, and trend views with novel interaction techniques, AirVis helps users uncover hidden patterns and relationships in air quality data, aiding in decision-making and scientific understanding.

What Problem Does This Paper Address?

Air pollution is a growing global concern, with significant health and societal impacts. While sensor networks are increasingly deployed to monitor air quality, the resulting datasets are often large, multi-dimensional (measuring various pollutants like PM2.5, NO2, SO2), and spatio-temporal. Analyzing these complex datasets to understand trends, identify anomalies, and discover underlying causes is a significant challenge. Existing visualization tools often struggle to handle the scale and dimensionality of this data, or they lack the integrated approach needed to connect spatial, temporal, and pollutant-specific information effectively. This paper aims to bridge this gap by providing a comprehensive visual analytics framework.

Methods

The core of AirVis is its integration of several visualization techniques and interaction methods:

  • Map-Based Views: Utilizes Google Maps to display station locations and overall air quality. It offers two metaphors: points representing individual station data and heatmaps for area-wide air quality. Interactions like polygon selection allow users to focus on specific regions or provinces.
  • Calendar View: A novel "Overview + Detail" calendar view is introduced. It visualizes daily air quality data in a calendar format, allowing users to grasp monthly, weekly, and daily averages. When zoomed in, it transitions to a rectangular view to show hourly patterns, revealing cyclical trends.
  • Trend View: Employs a Focus + Context approach with two connected line charts. The top chart shows the overall trend at a raw granularity, while the bottom chart displays a detailed view of a selected time period. A key innovation is a self-adaptive algorithm that automatically adjusts the granularity of both charts based on data size and temporal resolution, accommodating datasets with varying time granularities.
  • Multi-Dimensional View: For analyzing multiple pollutants simultaneously, AirVis uses a system where each pollutant is represented by its own map-based chart and calendar view. Trend views display multiple pollutants as separate lines on a single chart, enabling comparison of their temporal behaviors.
  • Adaptive Multi-Scale Trend View: This section details algorithms for determining appropriate time granularities for linear and cyclic analysis. It defines nine time levels (year, season, month, week, day, hour, minute, second, millisecond) and provides methods to automatically select the best granularity for overview and detail views based on data size and screen resolution.
  • Interactions: AirVis supports polygon selection on maps, brushing on trend charts to select time intervals, and tooltips to reveal detailed information. Clicking on map plots can trigger a detailed calendar view.

Datasets

The primary dataset used in this study is collected from air quality monitoring stations operated by the China National Environmental Monitoring Center (CNEMC). This dataset includes hourly measurements of pollutants such as SO2, NO2, CO, O3, PM2.5, and PM10 from 1437 stations across China. The data spans from January 1, 2014, to December 31, 2014, totaling approximately 12 million data items. The authors developed a web crawler to collect this hourly updated data.

Key Insights and Ideas

  • Integrated Visualization for Complex Data: The paper demonstrates the power of combining different visualization types (map, calendar, trend) to provide a holistic view of spatio-temporal air quality data, which is difficult to achieve with single-view approaches.
  • Adaptive Granularity for Time Series: The self-adaptive trend view is crucial for handling the diverse temporal resolutions of air quality data, making the system robust and user-friendly without manual configuration.
  • Connecting Spatial and Temporal Patterns: By linking map views with calendar and trend views, users can explore how air quality varies geographically and over time, revealing localized pollution hotspots and seasonal or daily cycles.
  • Uncovering Subtle Relationships: The multi-dimensional analysis, though not using traditional high-dimensional techniques, effectively highlights correlations and differences between various pollutants, such as the inverse relationship between PM2.5 and PM10 during certain wind events.
  • Domain Expert Validation: Positive feedback from domain experts validates the system's usability and its ability to help uncover interesting patterns that would be challenging to find with existing tools.

Key Results

The paper presents a functional visual analytics system, AirVis, and demonstrates its utility through case studies. Key qualitative results include:

  • Identifying that Northern China generally experiences higher PM2.5 levels than other regions.
  • Observing that PM2.5 concentrations are typically higher in spring and winter than in summer and autumn.
  • Discovering specific patterns in Beijing, such as a notable decrease in PM2.5 from February 13-28, 2014, with subsequent rises and falls, potentially linked to wind conditions.
  • Highlighting the complex interplay between PM2.5 and PM10, where wind can cause opposing trends in their concentrations.

Quantitative results are not the primary focus, but the system's ability to process and visualize a large dataset (12 million data items) in near real-time is implied.

Limitations and Future Work

  • Multi-Dimensional View Limitations: The authors acknowledge that their multi-dimensional view, while effective for a limited number of attributes (six in this case), doesn't employ traditional high-dimensional visualization techniques like parallel coordinates. They suggest evaluating these methods in future work.
  • Data Ingestion: The current system does not support interactively ingesting new data sources, which is a limitation for users wanting to analyze their own data.
  • Clutter in Multi-Dimensional Views: Domain experts noted potential clutter when analyzing many dimensions simultaneously.
  • Future Work:
    • Integrating other data sources (e.g., traffic, atmospheric variables, land use, economic data) to explore causal relationships of air pollution.
  • Enhancing the system to interactively ingest new data sources.
  • Further evaluation of traditional multi-dimensional visualization methods.
  • Investigating the impact of factors like wind patterns and temperature on air quality more deeply.
  • Potentially exploring the use of machine learning for pattern discovery within the visual analytics framework.

Summary updated

BibTeX
@inproceedings{du2017visual,
  author = {Yi Du and CuiXia Ma and Chao Wu and Xiaowei Xu and Yike Guo and Yuanchun Zhou and Jianhui Li},
  title = {{A Visual Analytics Approach for Station-Based Air Quality Data}},
  booktitle = {Sensors},
  year = {2017},
  url = {https://openreview.net/forum?id=RpOGuJ_knzF},
}
Figure