Mid-Infrared Spectrometer for Black Plastics Sorting Using a Broadband Uncooled Micro-Bolometer Array
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors
- The introduction provides a strong background on the significance of plastic sorting, current technologies, and the need for advancements in mid-IR spectroscopy. However, while it includes several references to regulatory frameworks, existing spectrometry methods, and past research, it would benefit from more discussion on prior works that attempted similar approaches. A clearer connection between past research and the proposed work would strengthen the introduction (state-of-the-art).
- The sentence in lines 21-23 should have a reference.
- Specify which type of plastics were used in the abstract section.
- Could you please place a reference for the paragraph from lines 87-95.
- The methods section provides detailed descriptions of the experimental setup, including the spectrometer design, optical components, and data processing pipeline. However, some details, such as the exact calibration procedure and steps for ensuring reproducibility, could be elaborated further.
- Could you give more details of the acquisition of the spectra in figure 8, such as resolution, number of scans. were these spectra averaged, do they have any pre-treatment ( smoothing or baseline correction)?
- Could you give more details of the acquisition of the spectra in figure 9, such as scan number, were these spectra averaged, do they have any pre-treatment (smoothing or baseline correction), what was the thickness of the plastic samples, and what is the distance between the IR source and the sample?
- Additionally, while PCA is used for classification, there is limited discussion. Do the spectra have any spectral pre-treatment in the PCA, such as Mean Centering or SNV? Can the authors give details of how many spectra were measured for each type of plastic? Do the spectra correspond to the same sample or different samples for the same type of plastic? What percentage of the data corresponded to the calibration set and the prediction set?
Can the authors specify which spectra correspond to the prediction and calibration set in Figure 11?
What software was used for PCA?
Author Response
The authors are thankful for the comments and suggestions of Reviewer 1, which allows us to improve the paper. We have made major revisions on the submitted version, including a substantial reorganization of the introduction part and state-of-the-art section; and clarifications on the data processing part, including a better description of the methods used. Please, find below, an answer to each comment. Note that, all numbering (figures, section, references) are from the previous submitted version.
[Comment 1] The introduction provides a strong background on the significance of plastic sorting, current technologies, and the need for advancements in mid-IR spectroscopy. However, while it includes several references to regulatory frameworks, existing spectrometry methods, and past research, it would benefit from more discussion on prior works that attempted similar approaches. A clearer connection between past research and the proposed work would strengthen the introduction (state-of-the-art).
[Response 1] We have reorganized substantially the introduction section, and added a state-of-the-art section, in order to facilitate the understanding the connections between the context, technological background, and this presented work. References [10, 11, 12] where the most similar approaches: but they couldn’t take advantage of a high sensitivity broadband infrared sensor as the one introduced in this work. In the revised version, we have made the connection more explicit between prior similar attempts and our work, while introducing the broadband micro-bolometer array.
[Comment 2] The sentence in lines 21-23 should have a reference.
[Response 2] We have added the following reference: https://doi.org/10.1016/j.resconrec.2022.106217
This a review article that concludes on the benefits of coupling infrared spectroscopy with machine learning for the recycling industry.
[Comment 3] Specify which type of plastics were used in the abstract section.
[Response 3] We have specified in the abstract the three plastic classes used in our simple classification study (PE, PET, PP).
[Comment 4] Could you please place a reference for the paragraph from lines 87-95.
[Response 4] We have added the following reference: Fourier Transform Infrared Spectrometry 2nd ed, by Peter R. Griffiths, James A. De Haseth. This is a reference book that describes in details the FTIR method.
[Comment 5] The methods section provides detailed descriptions of the experimental setup, including the spectrometer design, optical components, and data processing pipeline. However, some details, such as the exact calibration procedure and steps for ensuring reproducibility, could be elaborated further.
[Response 5] We have improved subsection 2.5: The calibration and pre-processing of camera frames is elaborated, using bullet points for more clarity. We gave more details on the calibration of the reference spectrum.
[Comment 6] Could you give more details of the acquisition of the spectra in figure 8, such as resolution, number of scans. were these spectra averaged, do they have any pre-treatment ( smoothing or baseline correction)?
[Response 6] As explained in subsection 2.5: “Unless specified, all spectra are extracted from 50 temporally averaged frames at 30Hz framerate, using the procedure described above without further post-processing.”. Moreover, we have added the settings used for the FTIR measurements (16 scans, 4cm-1 resolution).
[Comment 7] Could you give more details of the acquisition of the spectra in figure 9, such as scan number, were these spectra averaged, do they have any pre-treatment (smoothing or baseline correction), what was the thickness of the plastic samples, and what is the distance between the IR source and the sample?
[Response 7] The same response applied for spectra plotted in figure 9. All samples are deposited on the gold reference mirror, as described in figure 5. We have added more information on the thicknesses of the samples: there are all relatively thin samples of thicknesses from 50µm to 1mm.
[Comment 8] Additionally, while PCA is used for classification, there is limited discussion. Do the spectra have any spectral pre-treatment in the PCA, such as Mean Centering or SNV? Can the authors give details of how many spectra were measured for each type of plastic? Do the spectra correspond to the same sample or different samples for the same type of plastic? What percentage of the data corresponded to the calibration set and the prediction set?
Can the authors specify which spectra correspond to the prediction and calibration set in Figure 11?
What software was used for PCA?
[Response 8] Classification was not the main topic of the study, as we rely on simple machine learning methods, but we acknowledge that we need to provide more details on the methods used. Subsection 3.4 & Appendix A received significant clarifications, in the revised version. We use the scikit-learn library for Python, this PCA used the method of Halko et al. 2011. All spectra are standardized (centered and scaled by unit of variance). Only one spectrum is measured for each separate sample. Due to the small dataset (40 samples), validation is performed on a single spectra, excluded from the calibration dataset, and this process is repeated for every spectra (40 times).
Reviewer 2 Report
Comments and Suggestions for Authors
- Lines 26-29. The region between 900-1700 nm (0.9-1.7 micrometer) is usually referred as long-wave NIR, not short-wave NIR. Please revise. The SWNIR is from 700-1100 nm with Silicon used as detector.
- Lines 109-111. This part is not necessary. Please remove it.
- Equation 1 is not necessary. Please consider to remove it.
- Line 153. wavelength max=13 micrometer. Please revise.
- Line 284. Describe more on PCA. What is the algorithm used for PCA calculation? NIPALS?
- Figure 11. Use PC1, PC2, and PC3 as label in the x, y, and z axis (not PCA1, PCA2, PCA3).
- Line 298-305. It is unclear. How many samples used for testing? How to separate the samples into training and testing? PCA and SVM is different. PCA used all samples. For SVM, the samples must be separated into at least two sets: training and testing.
- The SVM method should be well described. There are many ways to calculate the SVM.
Comments on the Quality of English Language
It is recommended to check the English language.
Author Response
The authors are thankful for the comments and suggestions of Reviewer 2, which allows us to improve the paper. We have made major revisions on the submitted version, such as clarifications on the data processing part, including a better description of the methods used. We have also made a few edits to improve on the quality of the English language. Please, find below, an answer to each comment. Note that, all numbering (figures, section, references) are from the previous submitted version.
[Comment 1] Lines 26-29. The region between 900-1700 nm (0.9-1.7 micrometer) is usually referred as long-wave NIR, not short-wave NIR. Please revise. The SWNIR is from 700-1100 nm with Silicon used as detector.
[Response 1] We acknowledge that the denomination of the different infrared sub-bands is not homogeneous across the different fields of expertise. The term Short-Wave Infrared (SWIR) is commonly used by sensors and lens manufacturers to indicate the sensitivity range of the InGaAs photodiode (0.9-1.7µm). This term notably is used by the author’s affiliated company, Lynred, which manufactures InGaAs array sensors. In the revised version, we have added the term long-wave NIR as a possible denomination for this infrared sub-band.
[Comment 2] Lines 109-111. This part is not necessary. Please remove it.
[Response 2] We have removed this part in the revised version.
[Comment 3] Equation 1 is not necessary. Please consider to remove it.
[Response 3] We have removed Equation 1 in the revised version.
[Comment 4] Line 153. wavelength max=13 micrometer. Please revise.
[Response 4] The max wavelength of the spectrometer design is 13µm. This is a design choice based on the typical organic chemistry use-cases. This max wavelength is not to be confused with the max wavelength of the sensor, which is 14µm.
[Comment 5] Line 284. Describe more on PCA. What is the algorithm used for PCA calculation? NIPALS?
[Response 5] Classification was not the main topic of the study, as we rely on simple machine learning methods, but we acknowledge that we need to provide more details on the methods used. Subsection 3.4 received significant clarifications. We use the scikit-learn library for Python, this PCA used the method of Halko et al. 2011: a randomized truncated Singular Value Decomposition.
[Comment 6] Figure 11. Use PC1, PC2, and PC3 as label in the x, y, and z axis (not PCA1, PCA2, PCA3).
[Response 6] We have modified the figure 11 with the correct axis labels.
[Comment 7] Line 298-305. It is unclear. How many samples used for testing? How to separate the samples into training and testing? PCA and SVM is different. PCA used all samples. For SVM, the samples must be separated into at least two sets: training and testing.
[Response 7] Due to the small dataset (40 spectra / 40 separate samples), validation is performed on a single spectra, excluded from the calibration dataset, and this process is repeated for every spectra (40 times). We have clarified this part in the revised version.
[Comment 8] The SVM method should be well described. There are many ways to calculate the SVM.
[Response 8] We used the simple linear SVM. We added the following reference: https://doi.org/10.1007/BF00994018, which is the original article for linear SVM.
Reviewer 3 Report
Comments and Suggestions for Authors
The authors present the design, implementation, and testing of a mid-infrared spectrometer proof-of-concept. While the manuscript is interesting and valuable, I suggest the authors consider the following improvements to enhance its clarity and impact.
- It is not very common to include subsections in the introduction, and doing so has made this section more confusing rather than improving clarity. The authors begin the introduction well, but the addition of subsections (1.1, 1.2, and 1.3) disrupts the natural flow, making it harder to follow. I recommend revisiting the entire introduction and restructuring it to maintain coherence.
- What are the potential drawbacks of your system, and what limitations does this proof of concept have? A discussion on these aspects would strengthen the manuscript.
- How was the PCA conducted, and which statistical tool was used? The methods section lacks details on the statistical analysis, and this should be addressed.
- How do essential parameters such as spectral resolution, signal-to-noise ratio (SNR), and sensitivity measure up against current technologies?
Author Response
The authors are thankful for the comments and suggestions of Reviewer 3, which allows us to improve the paper. We have made major revisions on the submitted version, including a substantial reorganization of the introduction part and state-of-the-art section; clarifications on the data processing part, including a better description of the methods used; and better comparison of performances compared with existing methods. Please, find below, an answer to each comment. Note that, all numbering (figures, section, references) are from the previous submitted version.
The authors present the design, implementation, and testing of a mid-infrared spectrometer proof-of-concept. While the manuscript is interesting and valuable, I suggest the authors consider the following improvements to enhance its clarity and impact.
[Comment 1] It is not very common to include subsections in the introduction, and doing so has made this section more confusing rather than improving clarity. The authors begin the introduction well, but the addition of subsections (1.1, 1.2, and 1.3) disrupts the natural flow, making it harder to follow. I recommend revisiting the entire introduction and restructuring it to maintain coherence.
[Response 1] We have reorganized substantially the introduction section, and added a state-of-the-art section, in order to facilitate the understanding the connections between the context, technological background, and this presented work.
[Comment 2] What are the potential drawbacks of your system, and what limitations does this proof of concept have? A discussion on these aspects would strengthen the manuscript.
[Response 2] We have added a discussion about in subsection 2.1 about the main drawback of the technology, which is the thermal contrast time of the micro-bolometer of about 12ms. This effect is similar to a motion blur, that make the sensor unsuited for high-speed conveyor belts applications (>1m/s). We have added this drawback in the conclusion as well.
[Comment 3] How was the PCA conducted, and which statistical tool was used? The methods section lacks details on the statistical analysis, and this should be addressed.
[Response 3] Classification was not the main topic of the study, as we rely on simple machine learning methods, but we acknowledge that we need to provide more details on the methods used. Subsection 3.4 received significant clarifications, on the PCA, SVM and validation methods.
[Comment 4] How do essential parameters such as spectral resolution, signal-to-noise ratio (SNR), and sensitivity measure up against current technologies?
[Response 4] We have added a discussion about in subsection 2.1 about performance improvement compared with linear pyroelectric sensors on SNR (x244) and resolution (spectral sampling on 640 pixels against 128, 256 or 510 pixels). We have added these elements of comparison in the conclusion as well.
Reviewer 4 Report
Comments and Suggestions for Authors
The authors present a paper entitled “Mid-Infrared Spectrometer for Black Plastics Sorting using a Broadband Uncooled Micro-Bolometer Array”, in which they studied the design, implementation and test of a Mid-Infrared spectrometer, utilizing an uncooled micro-bolometer array. This instrument aids instantaneous measurements, comparable to a FTIR but with a more compact design and without moving parts. The Mid-IR range offers significant advantages over NIR-SWIR spectrometers especially for plastic waste sorting (also black parts). These materials were tested with significant accuracy and effectiveness and the plastic classification was performed with a very simple machine learning algorithm. Results of the paper are promising, and the dealt topic is a new one and it could be of interest for the readers. Results of the paper are well-presented and well-written and for this reason, I suggest accepting the paper in the present form.
Author Response
The authors are thankful for the kind comments of Reviewer 4.
Round 2
Reviewer 2 Report
Comments and Suggestions for Authors
This revised version is well-improved.
Author Response
[Comment from Reviewer 2] This revised version is well-improved.
[Response] Thank you for helping us improve the paper!
Reviewer 3 Report
Comments and Suggestions for Authors
The authors have incorporated the requested changes as indicated in their response document. However, it is difficult to locate these modifications in the manuscript since no color coding has been applied. Additionally, their response mentions revisions in Section 2.1 regarding the drawbacks, but I am unable to find these changes in the manuscript. Similarly, in Section 3.4, they reference PCA and SVM, yet these are not present in the manuscript. It is possible that there is a discrepancy in the subsection numbering. I kindly request that they highlight the changes for clarity. Thank you.
Author Response
[Comments from Reviewer 3] The authors have incorporated the requested changes as indicated in their response document. However, it is difficult to locate these modifications in the manuscript since no color coding has been applied. Additionally, their response mentions revisions in Section 2.1 regarding the drawbacks, but I am unable to find these changes in the manuscript. Similarly, in Section 3.4, they reference PCA and SVM, yet these are not present in the manuscript. It is possible that there is a discrepancy in the subsection numbering. I kindly request that they highlight the changes for clarity. Thank you.
[Response] The numbering was modified from the unrevised submitted version to the first revied version, due to the reorganization of the introduction and state-of-the-art parts. We will help you locate the requested changes, with the new numbering from the revised version. Please, see the main modifications highlighted in in the .pdf file attached.
Subsection 2.1 became subsection 3.1. the drawbacks of the proposed spectrometer is explained in lines [150-154]: “There is little interest to overclock the sensor up to higher framerates, due to the micro-bolometer’s thermal time constant τth ≃ 12 ms. This quantity characterizes the pixel’s response time to a fast temperature variation of the scene, resulting in a visual effect similar to a motion blur. This effect, inherent to thermal sensor technologies, makes such a sensor unsuited for high-speed conveyor belts applications (>1 m/s).” This drawback is also drawn in the conclusion [351-353]: “The main drawback of this technology, compared with a photo-diodes array, is that the pixel has a thermal response time of about 12 ms. This generates an effect similar to a motion blur, making it unsuited for high speed conveyor belts (>1 m/s).”
Subsection 3.4 became subsection 4.4. The PCA method is precised in line [299] and referenced with the article [24] Halko et al., 2011: “Finding Structure with Randomness: Probabilistic Algorithms for Constructing Approximate Matrix Decompositions” in SIAM Review
The linear SVM method is precised in line [318] and referenced in article [25] Cortes et. al., 1995: “Support-vector networks’’ in Machine Learning
Author Response File: 
Author Response.pdf
