Low-Cost Outdoor Air Quality Monitoring and Sensor Calibration: A Survey and Critical Analysis

Francesco Concas, Julien Mineraud, Eemil Lagerspetz, Samu Varjonen, Xiaoli Liu, Kai Puolamäki, Petteri Nurmi, Sasu Tarkoma

TL;DR

This paper surveys the rapidly growing field of low-cost outdoor air quality monitoring, focusing on the critical challenge of sensor calibration. It provides a comprehensive overview of sensor technologies, machine learning calibration techniques, and evaluation methods, highlighting open research challenges and future directions for more accurate and widespread air quality monitoring.

What Problem Does This Paper Address?

Air pollution is a major global health concern, and accurate monitoring is crucial for understanding and mitigating its effects. Traditional air quality monitoring stations are highly accurate but prohibitively expensive and sparsely deployed, leading to poor spatial resolution. Low-cost sensors offer a promising alternative for denser deployments, but their accuracy is often compromised by cross-sensitivities, environmental factors, and sensor drift. This paper addresses the need for robust and reliable calibration methods for these low-cost sensors to unlock their full potential for widespread air quality monitoring.

Methods

The paper provides a comprehensive survey of the current research landscape, categorizing methods and technologies into several key areas:

  • Sensor Technologies: It reviews various low-cost sensor types for gaseous pollutants (MOS, EC, NDIR, PID) and particulate matter (DiSC, LSP), detailing their advantages, disadvantages, and suitability for different applications.
  • Calibration Pipeline: The authors outline a typical machine learning pipeline for air quality sensing, encompassing per-device operations (design, placement, data collection, pre-processing, calibration, evaluation) and integration operations.
  • Machine Learning Calibration Algorithms: A significant portion of the paper is dedicated to surveying and critically analyzing various ML algorithms used for calibration, including:
  • Linear Models: Linear Regression (LR) and Multivariate Linear Regression (MLR) are discussed as baseline methods, noting their simplicity but limitations in handling complex relationships.
  • Support Vector Regression (SVR): Explored for its ability to model non-linear relationships.
  • Decision Trees and Random Forests (RFs): Discussed for their ability to handle non-linear problems and mitigate overfitting.
  • Boosting Algorithms: Including AdaBoost, Gradient Boosting, and XGBoost, for their sequential learning approach.
  • Artificial Neural Networks (ANNs): Including Feedforward Neural Networks (FFNNs), Recurrent Neural Networks (RNNs), Time Delay Neural Networks (TDNNs), and Nonlinear Autoregressive Exogenous Models (NARX), highlighting their power in modeling complex time-series data.
  • Gaussian Processes (GPs): Presented as a non-parametric approach with probabilistic outputs.
  • Advanced Paradigms: Transfer learning, semi-supervised learning, incremental learning, and dimensionality reduction are also covered.
  • Evaluation Metrics: The paper details various performance measures (MSE, RMSE, MAE, MBE, MRE, MAPE, R²) and evaluation criteria (Robustness, Resolution, Accuracy, Scalability) used to assess calibration methods.

Datasets

The paper does not introduce new datasets. Instead, it surveys existing studies that utilize various datasets, often proprietary and specific to individual deployments. The authors highlight a significant limitation: the lack of openly available, long-term, and spatially diverse reference datasets, which hinders objective comparison of calibration methods.

Key Insights and Ideas

  • The "Low-Cost" Trade-off: While low-cost sensors offer unprecedented spatial granularity, their inherent limitations (cross-sensitivities, drift, environmental influences) necessitate sophisticated calibration techniques.
  • Machine Learning is Key: Machine learning, particularly techniques that can learn complex non-linear relationships and temporal dependencies, is crucial for overcoming the limitations of low-cost sensors.
  • No One-Size-Fits-All: Different sensor types and pollutants require tailored calibration approaches. Developing universal models that perform well across multiple pollutants and sensor types remains a significant challenge.
  • Data is King (and Scarce): The performance of ML models is heavily dependent on the quality and quantity of training data. The lack of standardized, long-term, and diverse datasets is a major bottleneck for advancing the field.
  • Beyond Accuracy: While accuracy is paramount, other factors like robustness, resolution, and scalability are critical for real-world, long-term deployments.

Key Results

The paper is a survey and critical analysis, not an experimental study with novel results. However, it synthesizes findings from numerous studies, revealing:

  • Performance Varies: Different ML models show varying performance depending on the pollutant and sensor type. For instance, RF models often rank high in accuracy for many pollutants, while TDNN and NARX models show promise for NO2.
  • Sensor Technology Matters: MOS and EC sensors are highlighted as highly scalable due to their low cost and suitability for large-scale deployments, despite their calibration challenges.
  • Robustness is Lacking: Most studies use short test datasets, making it difficult to assess the long-term robustness of calibration methods. Datasets spanning over a year are rare.
  • Resolution is Improving: Many studies achieve high temporal resolution (e.g., 1-minute intervals), which is crucial for near real-time air quality services.

Limitations and Future Work

The authors explicitly identify several open research challenges and directions for future work:

  • Sensor Combination: Finding optimal combinations of sensors to capture multiple pollutants without compromising calibration performance.
  • Life Cycle Management: Developing methods for managing heterogeneous sensor deployments, including detecting sensor failures and minimizing maintenance.
  • Mobility Effects: Better accounting for the impact of sensor movement on measurement accuracy.
  • Universal Models: Developing models that perform well across multiple pollutants and sensor types.
  • Deep Learning Applications: Further exploration of deep learning techniques for calibration, addressing the need for large, diverse datasets and interpretability.
  • Dataset Standardization: The critical need for standardized, long-term, and spatially diverse datasets to enable robust model comparison and validation.
  • Concept Drift and Re-Calibration: Developing effective methods for detecting and mitigating concept drift, and implementing periodic or online re-calibration strategies.
  • Multiple Performance Measures: Moving beyond single performance metrics to a more holistic evaluation using multiple, complementary measures.
  • Virtual Sensors: Exploring the use of proxy variables to estimate pollutant concentrations, potentially reducing the number of required sensors.
  • 5G and Edge Computing: Leveraging advancements in networking and computing for real-time data processing and analysis in dense sensor networks.

Summary updated

Abstract

The significance of air pollution and the problems associated with it are fueling deployments of air quality monitoring stations worldwide. The most common approach for air quality monitoring is to rely on environmental monitoring stations, which unfortunately are very expensive both to acquire and to maintain. Hence, environmental monitoring stations are typically sparsely deployed, resulting in limited spatial resolution for measurements. Recently, low-cost air quality sensors have emerged as an alternative that can improve the granularity of monitoring. The use of low-cost air quality sensors, however, presents several challenges: They suffer from cross-sensitivities between different ambient pollutants; they can be affected by external factors, such as traffic, weather changes, and human behavior; and their accuracy degrades over time. Periodic re-calibration can improve the accuracy of low-cost sensors, particularly with machine-learning-based calibration, which has shown great promise due to its capability to calibrate sensors in-field. In this article, we survey the rapidly growing research landscape of low-cost sensor technologies for air quality monitoring and their calibration using machine learning techniques. We also identify open research challenges and present directions for future research.
BibTeX
@article{concas2021low,
  author = {Francesco Concas and Julien Mineraud and Eemil Lagerspetz and Samu Varjonen and Xiaoli Liu and Kai Puolamäki and Petteri Nurmi and Sasu Tarkoma},
  title = {{Low-Cost Outdoor Air Quality Monitoring and Sensor Calibration: A Survey and Critical Analysis}},
  journal = {ACM Trans. Sens. Networks},
  year = {2021},
  doi = {10.1145/3446005},
}
Figure