Evaluation and Calibration of Clarity Node S Low-Cost Sensors in Lubbock, Texas

Garber, John; Ardon-Dryer, Karin

doi:10.5194/egusphere-2025-4300

Preprints

https://doi.org/10.5194/egusphere-2025-4300

Preprints

23 Sep 2025

| 23 Sep 2025

Evaluation and Calibration of Clarity Node S Low-Cost Sensors in Lubbock, Texas

John Garber and Karin Ardon-Dryer

Abstract. Aerosol particles, also known as Particulate Matter (PM), have a profound impact on human health, air quality, the weather, and climate. PM can be measured using a variety of measuring techniques and instruments, notably reference-grade instruments and Low-Cost Sensors (LCS). Although Low-Cost Sensors allow for a higher resolution network, some have accuracy issues and reliability when compared to reference-grade units, which prompts the need to develop a calibration. This work, which is part of the Lubbock Environmental Action Plan (LEAP) for Communities, aims to provide information on air quality levels across the city of Lubbock using Clarity Node S sensors. In this study, which is the first step of the work, an evaluation and calibration of four Clarity Node S sensors was performed. The Clarity Node S sensors were selected for this project due to the sensors' ability to operate without a power or Wi-Fi source. Good agreement was found between the sensors when they were collocated with each other from March to May 2024 on the Aerosol Research Observation Station (AEROS). Next, one LEAP unit was collocated at AEROS with a reference unit, and different calibration tests were performed for the three PM concentrations measured by the Clarity units (PM₁, PM_2.5, and PM_10, particles with diameters <1, 2.5, and 10 µm, respectively). The selected calibration was developed and implemented for all four LEAP units. The calibrated LEAP units were then collocated near two different reference units for a duration of eight months (July 2024 to February 2025), and a comparison was performed. While one reference unit showed a good agreement with three LEAP units, the other reference units were very different from the collected LEAP unit.

Received: 02 Sep 2025 – Discussion started: 23 Sep 2025

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.

Download & links

Preprint (PDF, 1815 KB)

Supplement (733 KB)

Download & links

John Garber and Karin Ardon-Dryer

Status: closed

RC1:
'Comment on egusphere-2025-4300', Brayden Nilson, 23 Oct 2025
This study assess the performance of 4 Clarity Node S sensors collocated with/near an EDM and BAM FEM monitor for PM1, PM2.5, and PM10. They provide a good summary of their study site and instrumentation used, and a detailed description of their correction process. However, I fear their correction model is heavily overfit to the EDM data and needs to be re-evaluated. There is also a concern over the seemingly lack of pre-treatment (QA/QC) for the LCS data – it is well established that these monitors need to have erroneous values removed prior to correcting, typically through comparing the data from replicated internal sensors (ie. https://doi.org/10.5194/amt-15-3315-2022), and this likely explains the relatively low correlation of the raw PM2.5 data.
The authors do not mention splitting their data into testing/training sets like what Clarity Co. did for the correction they provided– without this there is a very high risk of overfitting, and the presented statistics will be biased. In addition, the EMD observations were included in the correction model, which will result in overfitting and a risk of the EDM observations dominating the corrected value. This is especially visible in Figure 7, where the corrected data nearly perfectly follows the EDM timeseries, in contrast to Figure 9 where the BAM monitor was used instead of the EDM. Due to the overfitting resulting from a lack of train/test splitting and the incestuous inclusion of the EDM data in the correction, I have serious concerns over the efficacy and transferability of the correction presented. The authors must split their data properly and seriously reconsider the inclusion of EDM data within the regression terms for this to be statistically sound.
I was expecting a discussion or conclusions section, but following the results there is just a summary section that repeats the methods and key findings. It could be helpful for the reader to have the large results section parsed into results/discussion/conclusion as is normally done.
See below for specific line-by-line comments and suggestions.
Specific comments:
Line 9: Use “LCS” acronym for “Low-Cost Sensors” once defined (i.e. “Although [LCS] allow for […]”)

Line 16: what is a “LEAP unit” – does this refer to the Clarity Node S sensors?

Line 20-21: specific numbers would be useful here, “were very different” is subjective

Line 25: suggested edit – “High concentrations of PM reduce air quality and produce negative impacts on human health”

Line 26: “Economy” should not be capitalized

Lines 33-62: suggest splitting this paragraph into two at essentially line 45; as a reader the transition from FRM to FEM felt unexpected, and readability would be improved with shorter paragraphs. You may need to add a third initial paragraph that introduces the terms FRM and FEM (basically expand on your first sentence of this paragraph) and potentially LCS as well, which would then flow nicely into the next 3 paragraphs, 1 for each of FRM, FEM, LCS.

Line 47: “and an optical monitor” feels like an after though – suggest expanding on to the same level of detail as the others

Line 62: capitalize “Low-Cost Sensors” to be consistent with the Abstract

Line 70: suggest finding a more up to date publication than 2015/16 given that the citations are used to say how there still are many uncertainties

Line 73: suggest adding a transition such as “In addition, LCSs can produce […]”

Lines 73-75: The “LCS” acronym is already plural, remove the “s” from “LCSs”

Line 76: For PM2.5, my experience has been very high correlation (>80%) with collocated FEMs post data cleaning: see table 1 of https://doi.org/10.5194/amt-15-3315-2022

Line 81: suggest adding the above publication as a citation as it builds on the Crilley et al method and compares with the Barkjohn et al correction

Line 96: suggested edit to remove duplicate usage of “climate”: “The [area has] a semi-arid climate, […]”

Line 124: edit for specificity: “[…] which are then converted to hourly and daily [mean averages] using MATLAB code”

Line 150: suggest removing “At the first step,” or providing a paragraph before this that gives a basic outline of each “step” so the reader has context

Line 155: clarify if the measurement every 15 minutes is an instantaneous sample or an integration over the 15 minute period

Line 170: combine these first two sentences into one, “Different calculations” is vague, and this sentence is essentially repeated on lines 170-171

Line 170-173: equations for metrics should be defined or cited

Line 173: remove capitalization of “Intercept”

Line 173: “best-fit information” is vague, clarify the regression method used (presumably linear regression based on the slop and intercept mentioned)

Line 176-195: suggest adding a table to display this and only referring to key numbers that help with the discussion. It is difficult as a reader to glean useful information from a long list of statistics.

Figure 2: font is small on the stats presented in each panel, requires zooming to 150%+ to be able to read – including a table like suggested previously could allow excluding this text and referencing the table instead

Lines 220-233: it makes sense that the correlation remained 66% (or slightly higher) after applying the LR and MLR given the nature of LR/MLR acting mainly to reduce bias. I have found from working with PurpleAir LCS that the best way to improve correlation is through QA/QC, not bias correction. QA/QC pretreatment was not mentioned in the methods, and could potentially significantly improve these results. Plantower sensors can report unrealistic (>2000 ug/m3) concentrations when sensors fail, or can have reduced sensitivity. These outliers can significantly impact correlation and bias/error. (you elude to this on line 263, but it could be worthwhile to expand the discussion of this). This also may explain the very low correlation observed for PM10 data on lines 366-370.

Line 245: RMSE is not a normalized metric, so it makes sense that a dramatic difference between studies can exist. It is possible that the Nobell et al (2023) study just had higher concentrations on average than this study.

Lines 248 and 256: “wondering” is something to be done in the discussion, to me this feels out of place in the results

Line 267: unclear what this means: “The MLR of this calibration was corrected using the following equation”

1 - 3: given that your calibration depends on observation data from the EDM monitor, how will this be transferable to other Clarity sensors? They would need to be colocated with an EDM to be able to apply this in real time, and if you have an EDM why would you setup a Clarity at the same location operationally? I have concerns about overfitting as a result of this as well, the improved performance relative to EDM could just be the result of the regression relying on the EDM data itself.

Line 287: “represents the interception” – this is vague, interception of what?

Figure 4: the colours used are difficult to differentiate especially with the dashed lines. Suggest splitting into 3 panels, each one comparing one of the LEAP timeseries with the EDM

Line 369: “the slope improved to 1” is not necessary to say, given that that is what LR does

Line 421: if the event was only for 1 hour, how were multiple EDM hourly mesaurements taken to produce the mean+/- SD?

Line 426: the fact that correcting the LCS made them detect the dust event when they did not before points to the overfitting resulting from including the EDM data in the correction

Lines 453-454: similar to the previous comment, the fact that after correcting the concentrations went from ~10 ug/m3 to ~100 ug/m3 shows how the EDM observations are dominating the corrected values – which is clearly visible to me in Figure 7

Lines 489-491: this reads a bit colloquially – specifically “puzzled” and “seem to be off”

Section 3.3.2: when switching to the BAM-1022 unit for comparison, was the same regression that was fit to the EDM monitor used? IF so, that would explain the poorer than expected performance given the over-fitting with the EDM data I have previously mentioned and the poor correlation between the BAM and EDM monitors you note on Line 562
Citation: https://doi.org/10.5194/egusphere-2025-4300-RC1
- AC1: 'Reply on RC1', Karin Ardon-Dryer, 13 Jan 2026
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-4300/egusphere-2025-4300-AC1-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2025-4300-AC1
RC2:
'Comment on egusphere-2025-4300', Anonymous Referee #2, 20 Nov 2025
The manuscript titled “Evaluation and Calibration of Clarity Node S Low-Cost Sensors in Lubbock, Texas” presents a rigorous and well-executed evaluation of four low-cost particulate matter (PM) sensors. The study addresses the critical need for affordable and dense air quality monitoring networks, particularly in regions lacking comprehensive regulatory coverage. Given the increasing relevance of low-cost sensors (LCS) in environmental and public health research, this work is timely and valuable. The methodology is robust, involving long-term collocation with reference instruments, regression-based calibration, and comparative assessments across multiple sites and time periods. The detailed description of experimental setup, calibration approaches, and statistical metrics (e.g., LR, MLR, RMSE, MAE, R²) enhances transparency and reproducibility. The introduction of PM₁₀ as a predictor variable is a noteworthy contribution, and the manuscript provides one of the first formal PM₁ calibrations for Clarity Node S sensors. Overall, the study offers important insights for deploying sensor networks in challenging environmental conditions and provides broadly applicable methodological lessons despite localized calibration results. The discussion of limitations is balanced and useful.
Limitations:
The calibration results are highly site-specific to the semi-arid, dust-influenced environment of Lubbock. While acknowledged, the generalization to other climates or aerosol regimes is not demonstrated.

The Clarity Node S PM₁₀ measurements show poor agreement with reference data, with regression failing to substantially improve accuracy. This limits the practical utility of the PM₁₀ channel in this context.

One of the reference instruments (BAM-1022) produced negative PM₂.₅ values in more than half the dataset. This complicates inter-comparisons and may introduce biases in calibration and evaluation.

Comments and Suggested Revisions:
(i) Some sentences are overly long or repetitive; tightening the prose would improve readability.
(ii) For several anomalies (e.g., July-August drift), the manuscript notes possible explanations but leaves them unresolved. More definitive interpretation or clearer acknowledgement of uncertainty would be helpful.
(iii) Lines 473, 561: Replace conversational phrasing (“We wonder…”) with objective academic language (e.g., “It was hypothesized that…”).
(iv) Throughout the manuscript, substitute informal expressions (“We were hoping…”, “We wonder…”) with impersonal scientific style.
(v) Define all abbreviations upon first use; terms such as “LT” and “TTU” are sometimes introduced without initial clarification.
(vi) Equations and variable formatting need greater consistency and clarity.
Citation: https://doi.org/10.5194/egusphere-2025-4300-RC2
- AC2: 'Reply on RC2', Karin Ardon-Dryer, 13 Jan 2026
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-4300/egusphere-2025-4300-AC2-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2025-4300-AC2

Status: closed

RC1:
'Comment on egusphere-2025-4300', Brayden Nilson, 23 Oct 2025
This study assess the performance of 4 Clarity Node S sensors collocated with/near an EDM and BAM FEM monitor for PM1, PM2.5, and PM10. They provide a good summary of their study site and instrumentation used, and a detailed description of their correction process. However, I fear their correction model is heavily overfit to the EDM data and needs to be re-evaluated. There is also a concern over the seemingly lack of pre-treatment (QA/QC) for the LCS data – it is well established that these monitors need to have erroneous values removed prior to correcting, typically through comparing the data from replicated internal sensors (ie. https://doi.org/10.5194/amt-15-3315-2022), and this likely explains the relatively low correlation of the raw PM2.5 data.
The authors do not mention splitting their data into testing/training sets like what Clarity Co. did for the correction they provided– without this there is a very high risk of overfitting, and the presented statistics will be biased. In addition, the EMD observations were included in the correction model, which will result in overfitting and a risk of the EDM observations dominating the corrected value. This is especially visible in Figure 7, where the corrected data nearly perfectly follows the EDM timeseries, in contrast to Figure 9 where the BAM monitor was used instead of the EDM. Due to the overfitting resulting from a lack of train/test splitting and the incestuous inclusion of the EDM data in the correction, I have serious concerns over the efficacy and transferability of the correction presented. The authors must split their data properly and seriously reconsider the inclusion of EDM data within the regression terms for this to be statistically sound.
I was expecting a discussion or conclusions section, but following the results there is just a summary section that repeats the methods and key findings. It could be helpful for the reader to have the large results section parsed into results/discussion/conclusion as is normally done.
See below for specific line-by-line comments and suggestions.
Specific comments:
Line 9: Use “LCS” acronym for “Low-Cost Sensors” once defined (i.e. “Although [LCS] allow for […]”)

Line 16: what is a “LEAP unit” – does this refer to the Clarity Node S sensors?

Line 20-21: specific numbers would be useful here, “were very different” is subjective

Line 25: suggested edit – “High concentrations of PM reduce air quality and produce negative impacts on human health”

Line 26: “Economy” should not be capitalized

Lines 33-62: suggest splitting this paragraph into two at essentially line 45; as a reader the transition from FRM to FEM felt unexpected, and readability would be improved with shorter paragraphs. You may need to add a third initial paragraph that introduces the terms FRM and FEM (basically expand on your first sentence of this paragraph) and potentially LCS as well, which would then flow nicely into the next 3 paragraphs, 1 for each of FRM, FEM, LCS.

Line 47: “and an optical monitor” feels like an after though – suggest expanding on to the same level of detail as the others

Line 62: capitalize “Low-Cost Sensors” to be consistent with the Abstract

Line 70: suggest finding a more up to date publication than 2015/16 given that the citations are used to say how there still are many uncertainties

Line 73: suggest adding a transition such as “In addition, LCSs can produce […]”

Lines 73-75: The “LCS” acronym is already plural, remove the “s” from “LCSs”

Line 76: For PM2.5, my experience has been very high correlation (>80%) with collocated FEMs post data cleaning: see table 1 of https://doi.org/10.5194/amt-15-3315-2022

Line 81: suggest adding the above publication as a citation as it builds on the Crilley et al method and compares with the Barkjohn et al correction

Line 96: suggested edit to remove duplicate usage of “climate”: “The [area has] a semi-arid climate, […]”

Line 124: edit for specificity: “[…] which are then converted to hourly and daily [mean averages] using MATLAB code”

Line 150: suggest removing “At the first step,” or providing a paragraph before this that gives a basic outline of each “step” so the reader has context

Line 155: clarify if the measurement every 15 minutes is an instantaneous sample or an integration over the 15 minute period

Line 170: combine these first two sentences into one, “Different calculations” is vague, and this sentence is essentially repeated on lines 170-171

Line 170-173: equations for metrics should be defined or cited

Line 173: remove capitalization of “Intercept”

Line 173: “best-fit information” is vague, clarify the regression method used (presumably linear regression based on the slop and intercept mentioned)

Line 176-195: suggest adding a table to display this and only referring to key numbers that help with the discussion. It is difficult as a reader to glean useful information from a long list of statistics.

Figure 2: font is small on the stats presented in each panel, requires zooming to 150%+ to be able to read – including a table like suggested previously could allow excluding this text and referencing the table instead

Lines 220-233: it makes sense that the correlation remained 66% (or slightly higher) after applying the LR and MLR given the nature of LR/MLR acting mainly to reduce bias. I have found from working with PurpleAir LCS that the best way to improve correlation is through QA/QC, not bias correction. QA/QC pretreatment was not mentioned in the methods, and could potentially significantly improve these results. Plantower sensors can report unrealistic (>2000 ug/m3) concentrations when sensors fail, or can have reduced sensitivity. These outliers can significantly impact correlation and bias/error. (you elude to this on line 263, but it could be worthwhile to expand the discussion of this). This also may explain the very low correlation observed for PM10 data on lines 366-370.

Line 245: RMSE is not a normalized metric, so it makes sense that a dramatic difference between studies can exist. It is possible that the Nobell et al (2023) study just had higher concentrations on average than this study.

Lines 248 and 256: “wondering” is something to be done in the discussion, to me this feels out of place in the results

Line 267: unclear what this means: “The MLR of this calibration was corrected using the following equation”

1 - 3: given that your calibration depends on observation data from the EDM monitor, how will this be transferable to other Clarity sensors? They would need to be colocated with an EDM to be able to apply this in real time, and if you have an EDM why would you setup a Clarity at the same location operationally? I have concerns about overfitting as a result of this as well, the improved performance relative to EDM could just be the result of the regression relying on the EDM data itself.

Line 287: “represents the interception” – this is vague, interception of what?

Figure 4: the colours used are difficult to differentiate especially with the dashed lines. Suggest splitting into 3 panels, each one comparing one of the LEAP timeseries with the EDM

Line 369: “the slope improved to 1” is not necessary to say, given that that is what LR does

Line 421: if the event was only for 1 hour, how were multiple EDM hourly mesaurements taken to produce the mean+/- SD?

Line 426: the fact that correcting the LCS made them detect the dust event when they did not before points to the overfitting resulting from including the EDM data in the correction

Lines 453-454: similar to the previous comment, the fact that after correcting the concentrations went from ~10 ug/m3 to ~100 ug/m3 shows how the EDM observations are dominating the corrected values – which is clearly visible to me in Figure 7

Lines 489-491: this reads a bit colloquially – specifically “puzzled” and “seem to be off”

Section 3.3.2: when switching to the BAM-1022 unit for comparison, was the same regression that was fit to the EDM monitor used? IF so, that would explain the poorer than expected performance given the over-fitting with the EDM data I have previously mentioned and the poor correlation between the BAM and EDM monitors you note on Line 562
Citation: https://doi.org/10.5194/egusphere-2025-4300-RC1
- AC1: 'Reply on RC1', Karin Ardon-Dryer, 13 Jan 2026
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-4300/egusphere-2025-4300-AC1-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2025-4300-AC1
RC2:
'Comment on egusphere-2025-4300', Anonymous Referee #2, 20 Nov 2025
The manuscript titled “Evaluation and Calibration of Clarity Node S Low-Cost Sensors in Lubbock, Texas” presents a rigorous and well-executed evaluation of four low-cost particulate matter (PM) sensors. The study addresses the critical need for affordable and dense air quality monitoring networks, particularly in regions lacking comprehensive regulatory coverage. Given the increasing relevance of low-cost sensors (LCS) in environmental and public health research, this work is timely and valuable. The methodology is robust, involving long-term collocation with reference instruments, regression-based calibration, and comparative assessments across multiple sites and time periods. The detailed description of experimental setup, calibration approaches, and statistical metrics (e.g., LR, MLR, RMSE, MAE, R²) enhances transparency and reproducibility. The introduction of PM₁₀ as a predictor variable is a noteworthy contribution, and the manuscript provides one of the first formal PM₁ calibrations for Clarity Node S sensors. Overall, the study offers important insights for deploying sensor networks in challenging environmental conditions and provides broadly applicable methodological lessons despite localized calibration results. The discussion of limitations is balanced and useful.
Limitations:
The calibration results are highly site-specific to the semi-arid, dust-influenced environment of Lubbock. While acknowledged, the generalization to other climates or aerosol regimes is not demonstrated.

The Clarity Node S PM₁₀ measurements show poor agreement with reference data, with regression failing to substantially improve accuracy. This limits the practical utility of the PM₁₀ channel in this context.

One of the reference instruments (BAM-1022) produced negative PM₂.₅ values in more than half the dataset. This complicates inter-comparisons and may introduce biases in calibration and evaluation.

Comments and Suggested Revisions:
(i) Some sentences are overly long or repetitive; tightening the prose would improve readability.
(ii) For several anomalies (e.g., July-August drift), the manuscript notes possible explanations but leaves them unresolved. More definitive interpretation or clearer acknowledgement of uncertainty would be helpful.
(iii) Lines 473, 561: Replace conversational phrasing (“We wonder…”) with objective academic language (e.g., “It was hypothesized that…”).
(iv) Throughout the manuscript, substitute informal expressions (“We were hoping…”, “We wonder…”) with impersonal scientific style.
(v) Define all abbreviations upon first use; terms such as “LT” and “TTU” are sometimes introduced without initial clarification.
(vi) Equations and variable formatting need greater consistency and clarity.
Citation: https://doi.org/10.5194/egusphere-2025-4300-RC2
- AC2: 'Reply on RC2', Karin Ardon-Dryer, 13 Jan 2026
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-4300/egusphere-2025-4300-AC2-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2025-4300-AC2

John Garber and Karin Ardon-Dryer

Supplement

https://doi.org/10.5194/egusphere-2025-4300-supplement

John Garber and Karin Ardon-Dryer

Viewed

Total article views: 7,180 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
5,113	1,885	182	7,180	290	139	187

HTML: 5,113
PDF: 1,885
XML: 182
Total: 7,180
Supplement: 290
BibTeX: 139
EndNote: 187

Views and downloads (calculated since 23 Sep 2025)

Month	HTML	PDF	XML	Total
Sep 2025	2,930	80	15	3,025
Oct 2025	550	100	30	680
Nov 2025	195	319	40	554
Dec 2025	276	327	25	628
Jan 2026	454	365	45	864
Feb 2026	274	309	14	597
Mar 2026	355	305	11	671
Apr 2026	78	67	2	147
May 2026	1	13	0	14

Cumulative views and downloads (calculated since 23 Sep 2025)

Month	HTML	PDF	XML	Total
Sep 2025	2,930	80	15	3,025
Oct 2025	550	100	30	680
Nov 2025	195	319	40	554
Dec 2025	276	327	25	628
Jan 2026	454	365	45	864
Feb 2026	274	309	14	597
Mar 2026	355	305	11	671
Apr 2026	78	67	2	147
May 2026	1	13	0	14

Viewed (geographical distribution)

Total article views: 7,184 (including HTML, PDF, and XML) Thereof 7,184 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 07 May 2026

Short summary

This study aimed to develop a calibration using a multivariate linear regression for the Clarity Node S for PM₁, PM_2.5, and PM₁₀. Calibrations for PM₁ and PM_2.5 were successfully implemented using internal temperature, relative humidity, and EDM-180 PM₁₀. Comparison of the calibration of sensors for eight months showed improvements in detecting spikes in high PM concentrations and maintained good correlation with a reference monitor.


Total:	0
HTML:	0
PDF:	0
XML:	0