A Machine-Learning-Based Study on All-Day Cloud Classification Using Himawari-8 Infrared Data

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors,

This manuscript introduces how to classify the cloud system based on infrared channels from HIMAWARI for all days. The authors derived ACCM model with XGBoost and then compared with the other machine learning models but there is problem in evaluation. That is why I recommend major revision before the publication.

 

Major Comments

1. Line 284-310, Table 7-8 and Figure 5: The authors compared the results from ACCM and the other ML models such as RF, LGBM. If you want to show the better performance of ACCM with XGBoost model, you should simulate the ACCM with RF, LGBM. I guess ACCM with RF or LGBM produces better results.

 

2. Table 3: As you know the ML technology, feature engineering is most important. That is why you show the Table 3 as a feature. But as an atmospheric scientists, you will have to explain the physical meaning of features. I mean why those BTS and BTDS are selected to classify the cloud pixels. BTD(3.9-11.2) is known as useful BTD to detect low level clouds. Like above, you will mention why those features are selected.

 

3. Classification is made for all days. Though, why are visible channels employed to classify? In particular, cumulus clouds is not easy to be classified due to their limited thickness. visible channels might be useful.

 

Minor Comment.

CPR/CALOP can be revised into CPR/CALIOP. 

There are so many typos related to CALOP. Then, CALIOP is name of sensor onboard CALIPSO satellite. You need to make it clear. CPR is onboard Cloudsat and CALIOP is onboard CALIPSO. 

 

 

 

Comments on the Quality of English Language

Revise typos.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

 

 

 

Review of A Machine Learning-Based Study on All-Day Cloud Classifica- 2 tion Using Himawari-8 Infrared Data 3 Yashuai Fu 1,2,†, Xiaofei Mi 3,†, Zhihua Han 1,2 , Wenhao Zhang 1,2,* , Qiyue Liu 1,2, Xingfa Gu 1,3, and Tao Yu 1.

The authors present an algorithm for cloud classification based on machine learning and using Himawari-8 data. The manuscript may be of interest, but there are some issues of methodology and discussion that need to be addressed. I found the body of the paper a bit confusing when it mentions that the data are collected during the day and night. Line 89: This study addresses the challenges associated with nighttime cloud classification and the limited number of classifiers by leveraging 10 IR, yet in line 133 the authors state that "This study focused primarily on cloud classification throughout the day, which requires the use 133 of 10 infrared bands, ranging from bands 7 to 16"....

 

I also recommend a more detailed description when the authors claim to use 10 channels but to integrate both datasets they use 9 cloud classifications (line: 163 To ensure consistency, the clouds were finally classified into 9 types based on one-to-one correspondence between the CPR/CALIOP and Himawar-8 clouds and line 184: The final set of input features included 5 BTDS, 10 infrared channels and latitude and longitude data). 

Equations 1,2,3 and 4 should show theoretical foundations with references.

In the body of the article, the authors state that "Line 130. 5 km spatial resolution but the table shows 2 km.

I understand that the method stated by the authors tries to better classify the clouds into different categories, but given that the authors have multiple clouds detected it would be relevant to show a more detailed description of the findings with the whole sample including the statistic effort considering all type, and showing the seasonal analysis. The isolated examples shown do not seem sufficient to validate the methodology. Moreover, it is not clear to what extent the improvement is a function of the number/type of clouds. In real terms, what is the effective improvement in, for example, daytime cloud classification?

The authors make a comparison with other methods that include daytime and nighttime clouds. It would be useful to make the comparison with other results that measure only daytime and only nighttime or including both scenarios, as otherwise they do not seem comparable.

 

Since this paper uses a single IR channel and assigns the cloud top temperature to the Tb value of this channel, the potential problems of this approach should be discussed further. Little mention is made of the fact that, being partially transparent to surface emission, high and thin clouds can produce higher Tb values than would be obtained from a cloud top correspondence with a temperature profile.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx