Deriving boundary layer height from aerosol lidar using  machine learning: KABL and ADABL algorithms

Rieutord, Thomas; Aubert, Sylvain; Machado, Tiago

doi:https://doi.org/10.5194/amt-14-4335-2021

Articles | Volume 14, issue 6

https://doi.org/10.5194/amt-14-4335-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

Special issue:

Tropospheric profiling (ISTP11) (AMT/ACP inter-journal SI)

https://doi.org/10.5194/amt-14-4335-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 14, issue 6

Research article

|

11 Jun 2021

Research article |

| 11 Jun 2021

Deriving boundary layer height from aerosol lidar using machine learning: KABL and ADABL algorithms

Thomas Rieutord, Sylvain Aubert, and Tiago Machado

Download

Final revised paper (published on 11 Jun 2021)
Preprint (discussion started on 07 Apr 2020)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of Rieutord’s “Mixing height derivation from aerosol lidar using machine learning: KABL and ADABL algorithms”', Anonymous Referee #1, 29 Apr 2020
- AC1: 'Response to Referee 1's comment', Thomas Rieutord, 30 Jun 2020
RC2: 'Review of article “Mixing height derivation from aerosol lidar using machine learning: KABL and ADABL algorithms”', Anton Sokolov, 05 May 2020
- AC2: 'Response to Referee 2's comment', Thomas Rieutord, 30 Jun 2020

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Thomas Rieutord on behalf of the Authors (06 Aug 2020) Author's response Manuscript

ED: Referee Nomination & Report Request started (28 Sep 2020) by E. J. O'Connor

RR by Anton Sokolov (18 Oct 2020)

Suggestions for revision or reasons for rejection

In the article “Mixing height derivation from aerosol lidar using machine learning: KABL and ADABL algorithms”, two machine-learning algorithms presented for the definition of the atmospheric boundary layer height, which is a principal parameter for the atmospheric modeling and air pollution dispersion.
I appreciate author answers to questions and suggestions of reviewers and I found that the second revision of the manuscript is definitely better than the first revision.
Nevertheless, I found that the structure of the article could still be improved. For example,
1. Line 223: “– classif_score: The internal score used to automatically choose the number of clusters (only used when n_clusters=‘auto’).” I suggest adding a reference to a following appropriate section.
2. Line 203: “1. Initialization: K centroids”. A question arises how this K is defined.

There are a few problems with equations and definitions:
3. Lines 270, 271 after equations, E and sigma are not defined; E is presented lately in line 312. Even if the expectation and the standard deviation are well known, they should be once defined in the manuscript.
4. Line 311: Please verify the equation as an expectation is a constant, and the variance of a constant is always zero.
5. Line 284: N and K are not defined in the equation

Other questions:
6. Line 188: “The accuracy was estimated by group K-fold, … “, do you mean K-fold cross-validation? In my opinion, the cross-validation is a key word.
7. Line 192: “An independent validation set was not used here because the accuracy was only used to discriminate between the classification algorithms.” Do you mean that the aim was to compare classification algorithms using cross-validation accuracy?
8. Line 252: “Finally, we look for the first change in the cluster attribution, starting from the ground level.” The justification of the algorithm should be somehow presented. I suggest adding an example of KABL classification as one as in previous author’s answer (question 21. Line 209) to illustrate the problem.
9. Line 434: “They occur with KABL because clusters do not always have vertical persistence (some points are identified as a free atmosphere in the middle of the boundary layer)”. As in the previous question, I suggest adding an example of KABL classification as one as in previous author’s answer (question 21. Line 209).
10. Line 301: “The most relevant configurations were retained and tested on the two-year dataset.” Do you mean the two-year radiosonde dataset?
11. Line 302: “There are height parameters…” , I believe that it is eight (8) parameters.

Hide

RR by Anonymous Referee #3 (03 Nov 2020)

Suggestions for revision or reasons for rejection

General comments:

The paper needs to undergo English proofreading.

For example, some small grammar mistakes exist throughout the paper, such as in line 72: “one of the six sensor”. Another example, the use of punctuation also requires to be improved. Please check the words: therefore, however, and so on.

Some detailed comments:

The motivation of using machine learning algorithms is not clearly indicated in the abstract.

Also, when performance was mentioned in the abstract (e.g. outperform, discrepancy etc.), the used metrics should be stated.

The abstract conclusion stating that ADABL is a promising algorithm, should be accompanied with key metrics evidences.

The motivations of the need of derivation of BLH needs to be improved. For example, Some sentences are not explained properly. For example, line 29: “Others methods are based on derivatives.”. The word of “derivatives” here are not clear, even after we read the next sentences. Another example, line: 34-35, the abbreviations of STRAT and CABAM are also not expanded and explained. Also, in line 38: “PathfinderTURB”.

The paragraph starting by line 40 should be used to explain AI first briefly, as some readers in journal may not be familiar with AI definition.

The authors need to describe clearly supervised and unsupervised learning concept, which can be done between lines and 55.

Some words are not known by wider audiences. For examples, line 95. The word row211 is not known well by the readers.

There are numerous supervised learning algorithms, the authors did not state clearly why boosting algorithms were chosen. Instead, the authors just stated that “the boosting algorithms are a very powerful family of algorithm”.

The KABL flowchart needs to be shown as a pseudo-code. The reviewer is also wondering, why the flowchart was made only for KABL, not for AdaBoost algorithm. If the paper attempts to compare both algorithms. They should be described fairly.

The notations are not fully described in equations (1) and (2).

The equations in subsections 3.4.2 are not numbered. The notations must also be explained properly. The reviewer understands that these are from textbook, but if the authors decided to include those in the manuscript. The notations must be clearly described.

Table 1 has no explanation. All of items inside the table must be described.

The fonts in Figure 9 are not readable.

Caption for Figure 11 needs to be expanded a bit.

Any idea, what is manufacturer algorithm does mathematically?

Please state future direction of the research and this work.

Hide

ED: Reconsider after major revisions (12 Nov 2020) by E. J. O'Connor

AR by Thomas Rieutord on behalf of the Authors (18 Dec 2020) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (12 Jan 2021) by E. J. O'Connor

RR by Anonymous Referee #3 (16 Mar 2021)

RR by Anonymous Referee #4 (18 Mar 2021)

Suggestions for revision or reasons for rejection

The manuscript presents a comparison of two machine learning algorithms for boundary layer height detection using MicroMPL measurements. The optimal configurations for the algorithms have been determined using a sensitivity analysis. The performances of both methods have been studied for a 2-year database and for a case study, comparing them with independent radiosonde estimates as well as with the estimates given by the instrument manufacturer.
The manuscript presents high scientific level, significance and potential for application of the work. In my opinion, the manuscript is well written, the different parts are clearly presented, and the approach is technically well justified and validated. The abstract is accurate and concise, the introduction is accessible, presents the topic background in a proper form and is followed by a clear presentation of the work done. Previous works on the subject are properly cited and the new points are clearly indicated. The methodology is properly explained, as it is expected in this kind of studies. The results and discussion sections are well structured and their content is relevant to properly analyze the proposed algorithms. Therefore, I propose that this article is accepted for publication, after improving some minor aspects that, to my view, will make the work more robust.
Main aspects:
-The concept of “mixing height” only appears in the title and in the conclusions. However, this concept is not properly discussed or explained, and the term “Boundary layer height (BLH)” is the one that is mentioned throughout the manuscript. I agree to use this more general name (BLH) in the present work, as I understand that the aim is not to separate or detect different phenomena within BL (e.g. mixing, stable nocturnal layers, etc.). Therefore, I think the term “mixing” should be avoided, in special for the title.
-To my opinion, the training of the manuscript using estimation “by hand” (as it is stated in line 170 of latest manuscript version, “On days where the boundary layer is easily visible to a human expert, the top of the boundary layer can be drawn by hand”) is well justified in the view of the explanations given in the following paragraph (lines 173-183). However, in order to avoid any doubt on the justification of this estimation, I recommend: 1) to include in the introduction some brief explanation with references on the relevant aspects of BL structure that are later mentioned in lines 173-183 (i.e. stable layer, etc) and 2) to give more quantitative information in such description (e.g. “a stable boundary layer is present near the ground during the night, as it is observed by the higher signal intensity due to higher aerosol concentration…” in line 175). These corrections will also help justifying other statements given in the manuscript, e.g. “The boundary layer was clearly visible and had nearly all the features of the conceptual image” in lines 413-414.
-Nighttime retrievals have been sometimes removed from the study, in particular for the overall comparison with RS (as it is stated in line 362) and for the monthly average (line 386). However, this is not well justified in any of the cases, and actually in section 2.2 it is said that the BLH was calculated using bulk Richardson number method for nighttime because it is a good method according to literature.
Other corrections:
-Lines 180-181: “The mixed layer started to develop at 09:00 UTC…”. In the view of Figure 4 (right), this development seems to start around 8 UTC actually.
-Both in Figure 4 and Figure 11, an indication on sunrise and sunset times would be helpful.
-Lines 190-192: lower case letters should be used after semicolon.
-Line 339: “optimal” instead of “most optimal”. This word already includes the superlative meaning.
-Caption of Figure 9: an indication of what INDUS means should be given.

Hide

ED: Publish subject to minor revisions (review by editor) (08 Apr 2021) by E. J. O'Connor

AR by Thomas Rieutord on behalf of the Authors (16 Apr 2021) Author's response Author's tracked changes Manuscript

ED: Publish as is (21 Apr 2021) by E. J. O'Connor

Short summary

This article describes two methods to estimate the height of the very first layer of the atmosphere. It is measured with aerosol lidars, and the two new methods are based on machine learning. Both are open source and available under free licenses. A sensitivity analysis and a 2-year evaluation against meteorological balloons were carried out. One method has a good agreement with balloons but is limited by training, and the other has less good agreement with balloons but is more flexible.