FRESH: An Autonomous IoT Platform for Multi-Parameter Environmental Sensing and Short-Term Forecasting

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents the IoT-based environmental sensing equipped with short-term forecasting. Even though the design and results sound good, several shortcomings should be improved:

1. The SOTA is limited. Therefore, the contributions and novelty of the work are not clearly shown.
2. The multi-parameter provides the benefits. The comprehensive experiments of these parameters show the main contribution of the work. However, too many parameters may confuse the user. Authors are suggested to address this issue.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. There is no complete information about the reference gas mixtures – in Table 3, "Not used" is indicated for NOx, VOC, O₃, CO, so₂ in the "Measurement Devices" column. How was the calibration accuracy assessed then? It is unclear whether verification was carried out after calibration for gas sensors other than temperature.
2. Calibration of NO₂ is not described sufficiently – it is claimed that a cylinder with a fixed concentration of no₂ was used, but the following are not specified: concentration value, exposure duration, number of repetitions, stability of conditions.
3. No separation into train/validation/test. It is described that the models were trained and evaluated on the August–October 2024 data. It is not specified whether a separate sample was used for testing. The reported R2 may be overestimated due to overfitting or using the same data for hyperparameter selection and estimation.
4. The choice of KNN (k=2) raises questions; for many variables, KNN with k=2 turned out to be optimal. This indicates a high sensitivity to noise and possible overfitting. The graphs of the error dependence on k for each parameter are not shown (only for PM2.5).
5. Lack of cross-validation – temporary cross-validation is standard practice for time series. The authors do not mention its use.
6. It is necessary to add a section where retraining issues will be discussed. It is necessary to provide arguments and data that will confirm that there is no retraining.
7. Low R2 for noise (0.071) is not explained. The authors state this fact, but do not offer explanations (non-stationarity, lack of patterns, lack of data). This is an important negative result that needs to be discussed in more depth.
8. It is necessary to align the tables exactly to the left edge. Now some tables have "moved out" to the right (for example, tables 1 and 2 have "moved out", and table 3 is located exactly).
9. It is necessary to set the same margins at the top of all tables (for example, table 2 has a larger margin at the top than table 3). The same applies to drawings – for example, Figure 1 has a larger indentation at the top than Figure 5.
10. What kind of picture is located between lines 264 and 265? The picture lacks a caption and numbering, and it is also unevenly positioned.
11. Figures 6, 7, and 8 are too small and need to be enlarged for better readability.
12. The margins between "Funding", "Conflicts of Interest", and "Abbreviations" are too large and of different sizes.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript presents an interesting IoT-based multi-parameter environmental monitoring platform with integrated short-term forecasting. The combination of heterogeneous sensors, GSM communication, local data logging, web dashboard, and battery-solar hybrid power supply is a strength, and the attempt to validate the system through real deployments is appreciated.
However, the manuscript still has important methodological weaknesses that limit confidence in the conclusions.
The main concern is the sensor calibration and validation. Although the manuscript refers to a “multi-point calibration procedure”, Table 3 suggests that several gas sensors were exposed to only one nominal gas concentration, which is not sufficient to support a robust linear calibration over the operating range. The supplementary material does not resolve this issue, as it mainly shows qualitative agreement in sensor response rather than a fully documented quantitative calibration.
In addition, Table 3 is internally inconsistent. The text states that certified gas cylinders were used as reference sources, yet the table reports “Not used” for several gas measurements. This should be corrected and clarified.
A second major issue concerns the thresholds used for exceedance events. Table 1 mixes different types of quantities and reference concepts, including physical concentrations, sensor index outputs, indoor limits, and 24 h average guideline values. As currently presented, the exceedance analysis is not methodologically clear enough.
Some of the reported environmental trends also require better explanation. In Figure 5, the apparent increase in PM exceedances during weekends is counterintuitive for a roadside location if traffic is assumed to be a dominant source, and UV should not systematically depend on weekday/weekend status. These results may reflect aggregation choices, missing data, or confounding factors, but they are not adequately discussed.
I also note some overstatement in interpretation. For example, the manuscript describes both correlations PM2.5–PM1 (0.83) and PM2.5–PM10 (0.51) as “strong positive correlations”, whereas 0.51 would usually be considered moderate rather than strong.
Figure 7 raises an additional concern about data completeness. Some site-specific weekly average series, especially for the Laboratory site, appear very sparse, suggesting incomplete temporal coverage. This should be explicitly reported, as such series should not be interpreted in the same way as more continuous ones.
The forecasting section is potentially interesting but still underdeveloped. Essential details are missing, including the exact train/test strategy, forecasting horizon, input features, and baseline comparison. The current results should therefore be interpreted more cautiously as a proof of concept.
Finally, although the manuscript claims solar-assisted autonomy as an important practical advantage, it does not quantify how much the solar panel actually extends battery life in field conditions.
In addition, the reference list should be carefully revised, as DOIs are missing for all cited papers.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have taken into account all the comments, and the work does not require any additional edits.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

I thank the authors for the careful revision of the manuscript and for the detailed response to the previous comments. The revised version is clearly improved, particularly in relation to calibration reporting, clarification of VOC/NOx index interpretation, explanation of data gaps, and the description of the forecasting workflow. These changes have strengthened the manuscript substantially.    
In my view, the paper is now close to being publishable, but a few final corrections are still needed.
- The manuscript should remain fully precise in distinguishing between quantitative calibration and functional verification, especially for the VOC sensors, so that the validation level of the different sensors is not overstated. 
- Table 1 should be checked once more carefully, as there are still minor clarity/presentation issues, including the duplicated Light entry. 
- Although the addition of DOIs is appreciated, the reference list still needs careful proofreading, since some entries appear inconsistent or potentially incorrect. For example, reference 16 appears to repeat the DOI of reference 15, reference 22 seems to report a DOI/link that does not match the cited journal/article details, and several entries still show formatting inconsistencies.  
- Some wording in the forecasting/results discussion should be made slightly more cautious, to remain consistent with the proof-of-concept nature of this part of the work.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf