A Machine Learning Approach to Derive Aerosol Properties from All-Sky Camera Imagery

Round 1

Reviewer 1 Report (New Reviewer)

The authors propose a machine learning-based methodology to derive Aerosol Optical Depth (AOD) and Angstrom Exponent (AE) from calibrated images of an all-sky camera. The approach establishes a relationship between AERONET measurements of AOD and AE and signals derived from the principal plane radiance measured by the camera at three RGB channels. Four models are used, differing in the input choice, and Gaussian Process Regression (GPR) is applied as the machine learning method. The novelty of the approach lies in obtaining a cloud-screened and smoothed signal that enhances the aerosol features contained in the principal plane radiance and can be applied in partially cloudy conditions. The authors suggest a quality assurance criterion that significantly improves their results, which show excellent agreement with AERONET measurements and have a high degree of numerical stability with negligible sensitivities to training data, atmospheric conditions, and instrumental issues. The authors' proposed methodology shows optimum performance and indicates that a well-calibrated all-sky camera can be routinely used to derive aerosol properties, making it ideal for aerosol research. Overall, the work may represent a significant contribution to aerosol monitoring.

The authors provided clear objectives for their research, however, there is an opportunity for improvement through additional revisions as suggested in the comments below.

 

 

Comments:

-          Scientific support in the form of relevant literature should be added to the phrases in this paragraph for better credibility. Specifically, additional references are needed for the statements made in lines 43-48, 63-65, and 83-86.

-          Lines 225-244: it is quite technical and complex, and may be difficult for some readers to understand. It would be helpful to provide additional explanation or context, or to break down the information into smaller, more easily digestible chunks.

-          Lines 396-405: The authors may have chosen a Gaussian Process (GP) with a Matern kernel of 3/2 as their machine learning model for their regression task because GP is a powerful and flexible non-parametric regression method that can provide probabilistic predictions with uncertainty estimates. The Matern kernel is a popular choice for GP because it can capture different levels of smoothness in the underlying function. Other regression methods such as linear regression, decision trees, random forests, or neural networks may also be suitable for regression tasks, but the choice of method depends on the specific problem and available data. I was wondering if you could provide a more detailed explanation of why you chose this specific model and kernel function over other machine learning methods.

-          Lines 511-515: It is recommended to add references for the used performance criteria  

-          The legends for Figures 2, 3, and 4 need to be improved for better clarity.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

In this manuscript, the authors present four methods to retrieve aerosol properties from an all-sky camera. Specifically, the Gaussian Process Regression is used to relate AOD and AE from AERONET measurements with signals based on the radiance of the principal plane. Experimental results show the predictions show an excellent agreement with the AERONET values. Generally speaking, this manuscript is well organized and interesting. However, there exist some points to be concerned.

 

Major Points:

 

1)     The first question is why the Gaussian Process (GP) regression  instead of other ML methods, e.g., Random Forest, Neural Network and so on, was chosen a tool to bridge the relationship between the aerosol properties, i.e. AOD and AE, and the AERONET measurements?

2)     It is necessary to compare the GP used in this manuscript with other ML methods.

 

Minor Points:

1)     There exist too many subfigures in Fig. 6, so that it occupies three pages.

2)     It is necessary to describe the evaluation metric, e.g., R2, RMSE. In addition, why the MAE was not used to evaluate the difference between the prediction and the ground-truth value? For example, as for the sentence “… for all the models and almost all predictions are close to real values within 0.02 23 for AOD and 0.2 for AE.”, you might use the MAE.

3)     Please rephrase the sentence in line 53 page 2.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report (New Reviewer)

Thank you for your efforts to address all of my concerns. I think that it is ready for publication now.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Some minor comments are listed below:

1. Line 125: Is "Cazorla et al. [" misplaced here?? 2. Authors are suggested to increase the fonts of some figures for better legibility for example Figure 1, 2, 3, 4, 6, 8 (axis lables).

3. Figure 4: There is something wrong with the y-axis labels.

4. Why are there 2 Conclusion sections (on page 17 and 19)??? May be the conclusion section is merged with Discussion section. Please correct this.

Reviewer 2 Report

1) I do not know reviewer 1 or reviewer 2, and you did not provide the complete questions.  I got there is one model now.  That model according to your text uses derived variables (e.g., AOD, AF).   

2) As for 'Why physically the less consuming time with this method' I can not say anything.  I can say that the one method used is flawed since the data upon which it is based are watered down by smoothing.  This is like rounding off the answer to a physics problem before the final step.  The model is not unique and predicts derived variables that can have same value but represent different underlying physics. Smoothing the data out lessens the scientific validity of the 'one method' and needs to be fixed, if not at the least clarified. Using AOT and AF might allow one to get away with the smoothing and the AI / ML technique might even indirectly replace some of the removed physics. But the uncertainty clouds the result.  Must use truly representative values of the phenomena you are trying to forecast or model to train the model.  If this was done, this reviewer missed this.  

 

3) Other illustrative example  contributing to decision. 

a. "...In the indirect 42 way (aerosol-cloud interactions) the aerosol particles may act as condensation nuclei for 43 the water vapor favoring the cloud formation that again reflect the Sun light tending to 44 cool the surface and warm it as greenhouse substance. Nowadays, we do not completely 45 understand the mechanisms behind these phenomena and the magnitude of the impact 46 of aerosols in climate remains highly uncertain [2][3]. I..."

            The first of these sentences  is not completely correct.  Must at least specify wavelength and cloud base altitude ... this is not presented here or noted.  Second sentence, Nowadays is not used in a scientific journal. This kind of language is possibly used in conversation or perhaps within a class.   

b. 'However, in complex situation or in boundary conditions, 81 so frequent in the real atmosphere, these assumptions may fail and then our physical 82 models too. On the other hand, all the power of fully learning approaches lies in the train- 83 ing data, since they do not use physical models. ".  

        The first of these sentences reads awkwardly. Your physical model is watered down and will fail under most situations. In the second sentence, yes this is true. Your training data needs to reflect true values of the variables being predicted. This is not the case based on the words read.  This marginalizes the credibility of the results.  

The English grammar was not improved.  It might even have been worse. I can speculate but will not.  

There are others throughout, however fixing these and indicating where in the revised text that occurs would help this reviewer.  

Reviewer 3 Report

This study proposed a ML modeling solution to derive the Aerosol Optical Depth (AOD) and Angstrom 10 Exponent (AE) from calibrated images of an all-sky camera. While the excellent results were approved by the experiment predictions through the main improvement on a clean, smoothed and cloud-screened signal able to highlight the aerosol features contained in the principal plane radiance, the modeling methods can be further tested under various meteorological conditions and air qualities. And this will ensure the reliability of the all-sky camera's capacity of estimating AOD and AE from images all the time. Meanwhile, more previous studies related to estimates of AOD and/or AE based on camara images need to be fully considered. In addition, principals or reasons for the development of four ML models are strongly suggested to add.