the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 07 Apr 2025
A Deep-learning Framework for Retrieving Tropical Cyclone Intensity and Structure from Gridded Climate Data (TCNN V1.0)
Abstract. This study presents a deep learning (DL) framework to retrieve tropical cyclone (TC) intensity and size from gridded climate data. Using a DL architecture based on convolutional neural networks (CNN) and the Modern-Era Retrospective analysis for Research and Applications (MERRA-2) reanalysis dataset, it is shown that our optimal CNN model for TC intensity retrieval (TCNN) can achieve a root mean squared error of 3–4 m s-1 at 0.5-degree resolution. With inherent constraints learned from the training data, the TCNN model can also retrieve the minimum central pressure and the radius of maximum wind with the mean squared errors of 10–12 hPa and 18–20 km, respectively, using the same unified model. Sensitivity analyses with different model configurations and input channels help identify the key factors and hyperparameters for TC intensity and structure retrieval in the MERRA-2 data. Examining the model performance using different data sampling methods reveals further that the TC information retrieval problem strongly depends on data sampling strategies. An improper sampling data could result in an overfitting of the model performance, which limits the application of DL models for downscaling or forecast purposes. Several potential improvements and challenges to handle this TC intensity data sampling will be also discussed.
- Preprint
(1266 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-1074', Anonymous Referee #1, 24 May 2025
This study presents a convolutional neural network (CNN) framework—TCNN v1.0—for retrieving key tropical cyclone (TC) intensity and structure metrics, such as maximum sustained wind speed (VMAX), minimum central pressure (PMIN), and radius of maximum wind (RMW), from gridded climate data. A major strength of this framework lies in its ability to infer realistic TC intensity characteristics from relatively coarse-resolution reanalysis (MERRA-2), addressing a long-standing challenge in global climate models where TC structures are typically under-resolved. The authors argue that this approach has the potential to improve TC intensity estimation from both current numerical weather predictions and future climate model outputs.
The study includes a thorough analysis of model sensitivity to input variables, domain configuration, and especially data sampling strategies. The results underscore the importance of proper train-test data partitioning, as the model’s performance degrades substantially when tested on unseen TCs using a chronological split. This finding is important and well-motivated. However, if the generalization issue is one of the study’s key conclusions, the decision to report the model’s primary performance metrics based on random sampling (where samples from the same TC may appear in both training and test sets) needs further justification. Specifically, while the reported RMSE for VMAX prediction (7.11 kt) appears to outperform previous methods, this result may overestimate the model's actual predictive capability, as the RMSE increases to 19.2 kt under a more realistic chronological split.
Furthermore, the authors cite existing studies such as Chen et al. (2019), which also employ CNN-based approaches to retrieve TC intensity from satellite data. Since Chen et al. used a chronological split in their validation, a more direct and critical comparison would be appropriate, even if the architectures and input data sources differ, especially given the common goal of improving TC intensity retrieval.
These issues also call into question the core assumption of the study—that ambient environmental conditions at 0.5° resolution contain sufficient information to estimate TC intensity. If the model struggles to generalize to new TCs, this may suggest that it is learning TC-specific patterns rather than robust physical relationships. As this assumption is foundational to the study’s broader claims, especially regarding the potential application to future climate projections, further justification or clarification is needed.
The sensitivity test on domain size (Section 3.2.1) is informative, and the conclusion that a 25°×25° input domain yields the best performance is reasonable. Still, more discussion linking the domain size results with those from model architecture and convolutional kernel experiments would strengthen the study. This would also help clarify how spatial context is encoded and used by the CNN. Similarly, the reported seasonal variation in TCNN performance deserves more physical interpretation, particularly regarding how environmental influences on TC intensity may vary by season.
In summary, while this study presents an innovative and potentially valuable approach for estimating TC intensity and structure from gridded climate data, the current manuscript does not yet provide sufficient justification for its core claims. The reliance on a data sampling strategy that inflates performance metrics, coupled with limited generalization to unseen TCs, raises concerns about the framework’s robustness and applicability, particularly for future climate projections, which inherently involve unseen conditions. Furthermore, the key physical assumptions underlying the model are not adequately supported by the results, and the sensitivity analyses, while informative, could be more cohesively interpreted to strengthen the physical insights.
Citation: https://doi.org/10.5194/egusphere-2025-1074-RC1 - AC2: 'Reply on RC1', Chanh Kieu, 19 Jul 2025
-
RC2: 'Comment on egusphere-2025-1074', Anonymous Referee #2, 17 Jun 2025
This manuscript, "A Deep-learning Framework for Retrieving Tropical Cyclone Intensity and Structure from Gridded Climate Data (TCNN V1.0)," presents the development and training of a convolutional neural network (CNN) model to retrieve tropical cyclone (TC) intensity, minimum central pressure (PMIN), and the radius of maximum wind (RMW) from the MERRA-2 reanalysis dataset. These quantities—intensity, PMIN, and RMW—are inherently inconsistent, challenging to observe directly, and typically poorly resolved in reanalysis products and climate models.
The authors aim to develop an enhanced tool capable of retrieving TC information from numerical weather prediction (NWP) models, with potential applications in future climate projections using gridded datasets. The novelty of this work lies in applying CNNs to coarse-resolution data (0.5° × 0.625°) to estimate point-scale TC properties, thereby bypassing the need for satellite imagery or high-resolution dynamic models.
Regarding model performance, the CNN demonstrates strong results. It is commendable that the model successfully retrieves VMAX values up to 155 knots. However, alternative configurations—such as different optimizers, loss functions, and filter arrangements—warrant consideration. Readers may also be interested in understanding the model’s comparative performance in single-output versus multi-output scenarios. Additionally, I recommend validating the multi output retrieval of PMIN and VMAX against empirically derived VMAX values based on PMIN to assess internal consistency.
The model is trained without differentiation across domains or seasons. It would be valuable to investigate whether dividing the model by domain or seasonal regime could reduce prediction errors. This is especially relevant since TCs exhibit basin-specific structural differences and seasonal variability.
Concerning the model domain, the sensitivity analysis helps justify the choice of an 18° × 18° domain size. The experiments span the North Atlantic (NA), Northwestern Pacific (WP), and Northeastern Pacific (EP), where this square domain is assumed to encompass both the inner and outer TC cores. However, it would strengthen the manuscript to explain the domain selection process in more detail. A schematic or graphical representation of domain placement relative to the TC center is highly recommended.
Another point of interest is the model’s ability to handle multiple cyclonic systems within a single domain. In active basins such as the NA and WP, it is not uncommon for two or more systems to exist simultaneously. Training the model to distinguish between multiple TCs and to identify respective parameters would be a useful enhancement.
The manuscript needs to thoroughly revised for the typos and scientific notations. For example, between lines 190 and 195, alternative to “data points” may be chosen for clarity. Similarly, there is a missing period at the end of the sentence on line 315.
Overall, this work represents a valuable contribution that could serve as a classifier for extracting TC information from reanalysis datasets. Looking ahead, the authors may consider extending the model to capture asymmetric TC structures. Since the best-track datasets include quadrant-specific wind radii (R1–R4) starting from 2014, future models could be trained to retrieve quadrant-based features rather than only RMW.
Citation: https://doi.org/10.5194/egusphere-2025-1074-RC2 - AC1: 'Reply on RC2', Chanh Kieu, 19 Jul 2025
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 834 | 86 | 18 | 938 | 16 | 23 |
- HTML: 834
- PDF: 86
- XML: 18
- Total: 938
- BibTeX: 16
- EndNote: 23
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1