Spatio-Temporal Alignment and Track-To-Velocity Module for Tropical Cyclone Forecast

Round 1

Reviewer 1 Report

The manuscript introduced a GRU-based neural network for tropical cyclone track forecasting. The Spatial Alignment Feature Fusion (SAFF) and Track-to-Velocity (T2V) modules are proposed to improve the previous works. The experiments on the historical North Indian Ocean TCs’ track dataset verified the effectiveness of the proposed method. Overall, the paper is interesting and sound. I only have a few concerns:

1.     Sections 3.5 to 3.7 of the experimental part should analyze the experimental results in more detail, instead of simply describing the experimental results.

2.     This paper only considers 24 h TC forecast. What about 48h, 72h even longer time span forecasts?

3.     Some recent weather forecast models, such as Pangu-Weahter, outperform the NWP method in TCs’ track forecast by direct forecasting the field data. In that case, please explain the advantages of the task in this paper.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

1. Ablation studies should be significantly extended and improved.

2. There are lot of missing details in the paper (e.g. what type of conv layers are used inside the model, better description of the model, maybe more visualisation of features)

3. References should be extended

4. Presented dataset is only one available?

5. Why Fusion Net results are only for 24h prediction, it is not possible to generate for other time periods?

6. Formatting should be improved, difficult to read, figures and tables appear earlier than they are cited in the article.

Formatting should be improved.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This is a nice paper that makes a significant contribution in the alignment of dynamic data and static data to better predict landfall of tropical storms. This is unique and highly useful especially with their stated 12 hour prediction window, which would be sufficient for most evacuations.  The extensive use of data (182 TCs and 2528 data points) is quite impressive

I think there are very few changes that are needed:

1. Move legend on Fig. 5 from side to top or bottom, which will let you make this set of figures much bigger. Please do so.

2. The literature background should be expanded to include a broader vision of matching of large remote sensing data sets

Please run Grammarly for a final check

Author Response

Please see the attachment.

Author Response File: Author Response.pdf