the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 26 Aug 2025
Enhancing Accuracy of Indoor Air Quality Sensors via Automated Machine Learning Calibration
Abstract. Indoor fine particles (PM2.5) exposure poses significant public health risks, prompting growing use of low-cost sensors for indoor air quality monitoring. However, maintaining data accuracy from these sensors is challenging, due to interference of environmental conditions, such as humidity, and instrument drift. Calibration is essential to ensure the accuracy of these sensors. This study introduces a novel automated machine learning (AutoML)-based calibration framework to enhance the reliability of low-cost indoor PM2.5 measurements. The multi-stage calibration framework connects low-cost field sensors to be deployed with intermediate drift-correction reference sensors and a reference-grade instrument, applying separate calibration models for low (clean air environment) and high (pollution events) concentration ranges. We evaluated the framework in a controlled indoor chamber using two different sensor models exposed to diverse indoor pollution sources under uncontrolled natural ambient conditions. The AutoML-driven calibration significantly improved sensor performance, achieving a strong correlation with reference measurements (R2>0.90) and substantially reducing error metrics (with root-mean-square error (RMSE) and mean absolute error (MAE) roughly halved relative to uncalibrated data). Bias was effectively minimised, yielding calibrated readings closely aligned with the reference instrument. These findings demonstrate that our calibration strategy can convert low-cost sensors into a more reliable tool for indoor air pollution monitoring. The improved data quality supports atmospheric science research by enabling more accurate indoor PM2.5 monitoring, and informs public health interventions and evaluation by facilitating better indoor exposure assessment.
Competing interests: Some authors are members of the editorial board of journal Atmospheric Measurement Techniques.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.- Preprint
(1624 KB) - Metadata XML
-
Supplement
(1628 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2025-3839', Anonymous Referee #1, 24 Sep 2025
-
RC2: 'Comment on egusphere-2025-3839', Anonymous Referee #2, 14 Nov 2025
This study presents a strong and timely contribution to improving the accuracy and reliability of low-cost indoor PM2.5 sensors, which are often affected by environmental factors like temperature and humidity, sensor drift, and nonlinear responses across pollution levels. The authors introduce an automated machine learning (AutoML)-based multi-stage calibration framework that connects low-cost sensors with intermediate drift-corrected references and a reference-grade instrument, applying separate calibration models for low and high concentration regimes. The methodology is clearly described, and the results are convincingly supported by statistical evidence. While the experimental campaign was limited in duration and conducted under specific semi-controlled chamber conditions, the proposed framework shows strong potential for broader application across different indoor (and outdoor) environments and long-term monitoring contexts. Overall, this is a well-written and well-executed study that offers a valuable and innovative advancement in automated calibration methods for low-cost air quality sensors, with the proposed multi-stage calibration framework significantly improving sensor performance.
I believe this study makes a valuable and well-executed contribution to the field, and I strongly recommend it for publication. The approach is innovative, the methodology is robust, and the results are clearly presented. I have only a few minor comments and suggestions listed below that should be addressed prior to publication.Specific comments:
- Line 146–152: Here, the authors describe which two types of low-cost sensors were used and against which reference instrument they were compared. I am wondering whether the low-cost sensors have a similar detectable particle size range as the Palas Fidas 200. It could add an additional layer of clarity to mention the particle size range for the reference instrument and - if available - also for the low-cost sensors.
- Line 154 – 155: Hopefully, the team got to enjoy the food afterward. Scientific dedication always deserves a good meal.
- Line 157: The phrase “natural indoor conditions” sounds somewhat contradictory, since indoor environments are by definition artificial. It might be clearer to use “realistic” or “typical”.
- Line 223–228: I wonder how the threshold of 50 µg m⁻³ was determined. The text mentions that the exploratory analysis revealed a bias flip at this concentration, but it could strengthen the explanation to briefly clarify why 50 µg m⁻³ was selected as the cutoff.
- Line 251: In the phrase “between 400-to-500 µg m⁻³” the hyphen is unnecessary. It would read more clearly as “between 400 and 500 µg m⁻³.”
- Figure 3: The two time series plots showing the raw data from multiple low-cost sensors are displayed as 3D graphics, which causes a slight misalignment between the sensor readings and the x–y axes, making the visualization somewhat confusing. I am not sure this is the most effective way to present the data, although it is not a major issue. I would generally recommend adding x-axis tick marks to indicate the days (with major ticks for the labeled days and minor ones for each individual day) and including grid lines especially in panels (a) and (b), which would help improve readability and reduce the visual confusion caused by the three-dimensional layout.
- Figures 4 and 5 appear to have different sizes and resolutions. Since both figures are quite similar, it would be visually more appealing and consistent to display them at the same size and resolution for better comparison and overall presentation quality.
- I really appreciate the thorough and transparent discussion of the limitations of the proposed method. The authors clearly acknowledge the constraints related to environmental conditions, emission sources, and long-term applicability. An analysis of long-term sensor drift would have provided valuable additional insights into the robustness of the calibration over extended periods. However, since this aspect is explicitly discussed as a limitation and identified as an important direction for future work, I find the current scope appropriate and well-justified for this study.
Citation: https://doi.org/10.5194/egusphere-2025-3839-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 1,944 | 78 | 12 | 2,034 | 37 | 45 | 35 |
- HTML: 1,944
- PDF: 78
- XML: 12
- Total: 2,034
- Supplement: 37
- BibTeX: 45
- EndNote: 35
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The paper proposes a novel AutoML-based multi-stage calibration framework to improve the reliability of low-cost PM2.5 sensors that are consistently influenced by sensor drift. The multi-stage calibration framework connects low-cost field sensors to be deployed with intermediate drift-correction reference sensors and a reference-grade instrument, applying separate calibration models for low and high concentration ranges. The proposed multi-stage calibration framework significantly improved sensor performance.
General comment:
The concept of the paper is timely and robust, and the authors provided clear explanation on the motivation and the calibration procedures itself. The results also look good as this multi-stage calibration seems to be able to help reduce errors. My main concern is how scalable this method can be. As the authors have also recognised the limitations that this relatively short period of measurements did not cover all emission source under all indoor environments. They suggested using the same method to cover more environments as future steps. To me, it is a well-written paper with clear concepts and good results. I would recommend for publication after addressing my minor specific and technical comments (listed below).
Specific comments:
Technical comments: