AICCA: AI-Driven Cloud Classification Atlas

Round 1

Reviewer 1 Report

The authors generate cloud cluster using only raw multispectral imagery as input, providing a really interesting approach of classifying cloud patterns, without the need of apriori cloud classification definitions. However, I am concerned about the suitability of the manuscript to Remote Sensing, as in its current form it consists an algorithm description, with many code details. Apart from that, any robustness of the results cannot be concluded, before the algorithm is being applied in a wider dataset over different geographical regions. Moreover, the script code plots are better to be removed in the Appendix part. Therefore, I accept the manuscript after some minor corrections and  explanations. Authors must consider the general and specific comments reported below.

Broad comments

Line 133-134: “Label is an integer in the range 1..42“. How was the range between 1 and 42 was selected? How was configured for 42 clusters?

How are the optimal defined clusters by the algorithm assigned to common cloud types? Thin, thick, cumulus etc (Cu, Sc, Ac, As)..

You mentioned that “suggesting that too many clusters may harm understanding of cloud patterns by excessive differentiation of fine detail. “. Could be the same case in your analysis? How are these clusters assigned to clouds with significant different physical properties?

Specific comments

Line 222: “Because do not expect our RI autoencoder to be robust to the selection of MODIS data “. Please rephrase, I guess you mean “Because we do not expect “…

Line 443: “These are are shown as the dots in Figure“. Please delete the second “are“.

Figure 5, 6 and 7. What are the units of the cloud optical thickness.

 

Author Response

"Please see the attachment."

Author Response File: Author Response.docx

Reviewer 2 Report

 

Summary:

Clouds are a major variable for climate modeling and climate projection. Traditional cloud classification schemes are simplistic and tedious to apply manually. They may also not capture the full diversity of cloud types and properties relevant to climate modeling. The authors explore unsupervised cloud classification on a massive dataset to overcome these issues. The authors process two decades of global MODIS multispectral satellite imagery and derive 42 cloud types with an unsupervised clustering approach. These classes form the AICC – an AI-driven Cloud Classification Atlas. The authors argue that the new cloud atlas consists of reasonable classes and enables a new data-driven perspective of cloud analysis for climate research.            

 

The unsupervised clustering approach follows four steps corresponding to the common practice: Data preprocessing, dimensionality reduction, clustering, and cluster interpretation. The major challenges of any clustering approach are the adequate choice of the cluster numbers and the interpretation of how the results relate to the phenomenon under study. For dimensionality reduction, the authors choose an autoencoder with a loss function that provides rotational invariance. Clustering is performed with the hierarchical agglomerative clustering (HAC) algorithm. The authors apply a well-thought-out strategy to determine the number of clusters. They successfully check for cluster stability, cluster significance, and intra-class-distance, which yields a reasonable cluster number of 40<k<48. The final choice is k=42 based on the self-defined WASD-parameter. The clusters are reasonably stable across winter and summer seasons and generally agree with the previous classification schemes and physical properties such as COT, COP, and ice fraction.               

 

Further, the authors carry out two case studies. The classification scheme is applied to two regions in the southeastern Pacific. The global spatial distribution of cluster regions is a remarkable result. Besides the mere cloud types in traditional classification schemes, the new cloud classes also contain information about the spatio-temporal occurrence of the clouds associated with more complex phenomena such as tropical convection or closed cell/open cell clouds. Both case studies suggest that the new classes contain richer information about clouds as traditional classification schemes.

 

General points:   

I would like to congratulate the authors for submitting this manuscript. It is an interesting study to explore the underlying patterns of clouds from such a comprehensive dataset. This study may open up new research directions to better understand the interaction between clouds and the climate. However, I have some minor concerns before this paper is ready for publication.

           

The abstract, the introduction, and the conclusion are already good but might benefit from a mild revision in which you check for consistency and a more concise presentation of your results. For example:

 

I think the wording is a little vague and only lists the results. Please revise the abstract (especially lines 12-17). 

 

The manuscript mentions the improvement of climate research as the overarching motivation in many places. However, the introduction only refers to the "needs of climate research." Please explain this point in more detail. 

 

The case study of the global distribution of cloud patterns is, in my view, an important result. This should become more credited in the conclusion and the abstract. Lines 602-603 read: "We conclude that (1) our methodology has explanatory power, in that it captures regionally unique cloud classes, and (2) 42 clusters is a useful number for a global analysis." This sentence is concise and should be placed in the conclusion and the abstract. 

 

Further, cluster stability and the relationship to the ISCCP are not mentioned in the abstract but play a major role in the study. Please put this in the abstract.

 

All in all: please double-check the abstract, the introduction, and the conclusion

to provide a concise overarching line of argumentation consistent with the rest (Section 2-6).

 

Detailed comments by line number:

188-190: please consider writing 1), 2), 3) as elsewhere in the manuscript 

 

240-241 Equation 1: Because the autoencoder takes multiple parameters, theta should be a

vector. Typesetting theta in bold also prevents confusion with a rotation angle because theta is often used as such.

 

252-253 Equation 4: How is the rotation operator R defined? How many rotations do you perform? Which angles do you use?

 

263: Why are the numbers for lambda_inv and lambda_res so high? If you minimize a function and weight the components, only the relative values should play a role?

 

 267-269: How did you check that 100 epochs are adequate and neither under nor over-fitting occurs?

 

547 – 548: I think this sentence makes no sense: It will be clear that the clustering algorithm will generate a unique geographical pattern if the input is a unique geographical cloud pattern in the satellite image. The wording "whether" is off. I would suggest just writing, "we investigate the nature and distribution..."

 

589: Clusters 21 – 22 appear to dominate in the north polar region and a band in the south. Could you explain this?

 

Figure 9: Surtitles à Subtitles

 

694: stale --> stable

 

 701: geographic --> global

 

Author Response

"Please see the attachment."

Author Response File: Author Response.docx