General comments:

This paper focuses on the identification of TCs based on satellite cloud images via DCNN techniques. Two models are proposed to deal with identification issues associated with two kinds of SCIs that are widely utilized in this field. Visualization techniques are further adopted to examine how the DCNN models work internally. Overall, the article is well organized and written. Both the methodology (including datasets and models/methods) and main results are presented and discussed clearly and reasonably. The results are interesting and useful. This reviewer suggests the article be accepted after minor revision.

Specific Comments:

1. Abstract: “lack of concerns on the identification of TC fingerprint from SCIs have become a potential issue, since it is a prerequisite step for follow-up analyses”. Please revise this sentence to improve its readability, meanwhile, have may be replaced by has.
2. L66: there lacks a blank
3. Line 131: tend--tends
4. Section 2.2.1: the authors use rotation technique for data augmentation. As discussed in this section, some information of the image may be lost. Will this operation result in any influence on the identification results?
5. Line 148: have—has
6. Line 174: to judging
7. Lines 178, 182, 208: format (especially for where)
8. The authors proposed two DCNN models. Although associatedprediction results seem to be quite good, how about the comparative performance of these models against others?
9. The authors report two types of heat maps which vary with each other evidently. Are there any reasons for why there will be such two kinds of heat maps?
10. Section 3.2.1: it seems that to use the IP technology the authors have to extract zoom-in view of TCs from the NWPO picture If it is the case, how to do this?

General comments:

In this study, a deep convolutional neural network (DCNN) is adopted to identify TC satellite images. Efforts are also made to explore how the DCNN models work internally. Overall, the work is interesting and the manuscript is well written, with analyses and discussions presented comprehensively and reasonably. I think this is a good piece of work that contributes to TC studies. I have only some minor comments which are given below. It is suggested the article be accepted after minor revision.

Specific Comments:

1. Line 24: ever—every
2. The authors claim that the normalization of image pixel values can accelerate the convergence of the model. Why?
3. Please explain a little more about the Dropout layer
4. Results of the evolutional curve in Figure 5(b) suggest that the training accuracy is not improved consistently with the increase of training epochs. When will you stop the training process then?
5. The authors present many heat maps of TC images. How about those of non-TC images? Are there any typical differences between these two types of heat maps?

This paper presents a research on classification of TC and non-TC pictures from satellite cloud images using the Deep convolutional neural network. Two image sets are used: the image set that covers all the Northwest Pacific Ocean basin with multi TCs, and L image set that covers small region of NWPO with single TC. The images are break out into training, validation and test sets.

Two DCNN models are trained for the two sets respectively. For the model trained with larger size images with multi TCs, the image pyramid technique is used to pre-process the images before training the model. Appropriate performance parameters are employed to evaluate the adequacy of the models. It shows that the pyramid technique improves the accuracy of the model.

The structures of the DCNN are well designed and presented in the paper. The results are well analyzed with proper discussion. The findings in this research should be valuable for further researches on this aspect, and even the models could be a useful basis for the meteorological agents to build their operational model on. 

Some questions and minor suggestions:

• Line 190：“Based previous tests” should be “Based on previous tests”
• Below section head 3.1.1：What does “The 10-fold cross-validation” mean? How is the “10-fold cross-validation” operated?
• In Figure 5, what is the difference for TG-1 to TG-10? Any difference in parameter setting among them? The same question for TG-1 to TG-3 in Figure 14.