Review of Wang et al. “Forecasting tropical cyclone tracks in the Northwest Pacific based on a deep-learning model”

In this work, Wang et al. use various deep-learning methods to forecast tropical cyclone tracks in the Northwest Pacific. The RNN, LSTM, and GRU models are used with historical trajectory data only; the best performing GRU model is then combined with a convolutional neural network (CNN) to capture time-dependent characteristics for improved accuracy. The model presented in this work has good performance (average distance error of 17.22km and 43.90km at 6h and 12h forecast) compared with the Central Meteorological Observatory’s results in short-term forecasting (27.57km and 59.09km, respectively), and can give forecast results in seconds.

Despite the promising nature of this work, which is in scope for Geoscientific Model Development, I have some concerns about the presentation and writing of this manuscript that I recommend careful revisions before further consideration.

Major comments:

1.    The manuscript is lengthy at 28 pages, but nearly half are devoted to introduction and methods that use a lot of space to describe standard procedures in machine learning. For example, section 2.2.1 goes into great detail about OOB samples. A section in feature selection using random forest is necessary (Figure 1 is very good) but the background is very standard, and can be shortened and refer the reader to appropriate background references. Section 2.2.2, 2.2.3 include a lot of background in RNNs and CNNs that can also be shortened as it is standard procedure in machine learning and not specific to forecasting cyclone tracks. Same for 3.3, 3.4 in normalization and evaluation criteria which do not need to be in separate sections; overall, the sections before the results can be reorganized for conciseness and avoid replicating a lot of existing background literature on the topic.

2.    I would like to suggest authors to be more careful in the introduction in regards to the strengths and limitations of NWP, statistical models, and deep learning, avoiding potential biases towards methods that are not deep learning. For example, line 49-50 says that NWP models have “limitations in methods” requiring “numerous calculations”. Framing of these limitations is needed. Is computational performance of these numerous calculations unacceptable? The work presented in this manuscript is of course efficient, giving results in seconds. But how long do NWP models take? The authors then mention “accurate mathematical descriptions of physical atmospheric mechanisms”. NWP models aren’t always exact and can involve many approximations and parameterizations. The authors may be trying to convey that NWP models require description of physical processes versus machine learning methods that learn from data. But that is not a limitation, it would be a property of different approaches to modeling.

The authors also criticize statistical models, saying “manual feature selection is unable to produce accurate predictions”. The inaccuracy would need to be characterized (and cited where appropriate) in order to reach this conclusion.

In the following paragraph about deep learning, authors give very specific accuracies (e.g., L83, L85, L91) and strengths of this method. For completeness and fair comparison, I suggest authors also give conventional methods similar statistics and strengths, and avoid vague, uncited description of limitations.

There are also other statements in the introduction that require revision. I include them in the specific comments below.

3.    Text in figures is at times hard to read because of the small font size. Please also make the fonts consistent, e.g., Arial. I suggest going through the futures to ensure consistency in presentation and that all figures are clear and readable.

Specific comments:

1.    Abstract, L29: please include the average distance errors of the CMO forecast results as well for comparison.

2.    L82 uses 24-h prediction distance error for LSTM, then L85 uses 6-h RMSE, L91 uses 6-h distance error. If possible, please be more consistent in error metrics.

3.    L86: authors say Kim et al. (2018) are “significantly better” than those of a CNN. How much?

4.    L97: please define MLP, first time the acronym has appeared in the manuscript.

5.    L101: “Previous studies have shown…”. Which previous studies? Please provide references.

6.    L102: “Still, most of them have neglected to describe and analyze the meteorological factors that affect the movement of TCs, ignoring valuable features.” Which studies? Did this neglect of meteorological factors significantly affect performance, compared to studies that have considered these factors? Please also give examples of these “valuable features”.

7.    L126-127: Do you mean that the Coriolis parameter is included in the predictors?

8.    L128-133 describes a TC bias to northwest; I am having trouble following the reason for this paragraph. Is this the reason for the geographically asymmetrical data selection in L142-147 (3)? If yes, then why is this not done for (1) & (2)? 

9.    L143, L145 “10 degree radius”. Do you mean extended by a 10 degree distance in each direction, since a 21x21 grid is formed?

10.    Figure 4: I suggest also adding the RMSE for the test set for the three recurrent neural networks inset in the figures for ease of comparison.

11.    L414, Figure 9 legend: 2106 -> 2016.

12.    L416-417: add “using deep learning methods” at the end of the opening sentence.

Review of Wang et al. “Forecasting tropical cyclone tracks in the Northwest Pacific based on a deep-learning model”

In this work, Wang et al. uses various recurrent deep learning models (RNN, LSTM, and GRU) with historical trajectory data to forecast tropical cyclones track in the Northwest Pacific. To further improve the forecast accuracy, the authors incorporate meteorological data by the use of CNN models. The combined model GRU_CNN outperforms other recurrent models and the results of Central Meteorological Observatory.

Overall, the results of the work look promising. However, I think the manuscript needs major revision before further consideration.

General comments:

1. The authors spend a lot of space explaining the technical details of the standard machine learning and deep learning models (random forests, RNN, CNN). These can be shortened, or refer the readers to the detailed background references. In addition, the equations for the GRU and LSTM cells are hard to follow, thus they can be complemented with diagrams showing the flow of data in these cells.
2. In the model framework section, I find it hard to understand the network architecture that the authors used in this work.
1. I think it would be benefit to include a table detailing the network architecture.
2. In addition, in figure 3, I think the description of the figure could be revised to include more details such as: CNN kernel size, what the solid white arrows mean, what the dashed red arrow means, etc.
3. The authors do not mention the architecture of the RNN, LSTM, GRU that they used in this work. I think it would improve the clarity if they were included here.
3. In the discussion of table 3 (L384-L392), the authors claim that the influence of SST and geopotential height gradually increases at long-term forecasts. Can the authors provide more explanation of why this is the case?
4. Since the authors compare the performance of GRU_CNN with other methods: FAXAI, MITAG, and IN-FA in figures 7-9, I think it would be more convincing if the authors can also provide detailed comparison between these models like in the table 3.

Specific comments:

• L79: missing a space between a reference and the word “applied”
• Figure 4: this figure could instead show the difference between the predicted longitudes/latitudes with the observed longitudes/latitudes to improve clarity and readability.
• All figures’ texts and labels can be a bit bigger to improve readability.
• L397: what are these methods: FAXAI, MITAG, IN-FA? Can you provide a short description and references for these methods?