Deep Learning Imputation of Missing Radius of Maximum Winds (Rmax) Values in Tropical Cyclone Best-Track Data

This study evaluates data-driven strategies for imputing missing radius of maximum winds (Rmax) values in tropical cyclone best-track data, comparing temporal deep learning models (one-dimensional Convolutional Neural Networks, 1DCNNs, and Long Short-Term Memory networks, LSTMs) against non-temporal machine learning models (XGBoost and artificial neural networks, ANNs), as well as the CK22 physics-based parametric model. The research uses observational data from the International Best Track Archive for Climate Stewardship (IBTrACS) and explores three input configurations: with R34, without R34, and a physics-augmented configuration based on Chavas and Knaff (2022) that includes additional derived parameters such as R342, Coriolis parameter, sin(lat), and the ratio Mmax/M34.

The results demonstrate that including the radius of 34-knot winds (R34) substantially improves performance across all model types. Among the evaluated models, the physics-augmented LSTM achieved the highest average correlation and the lowest average RMSE (0.781 correlation and 12.66 nm RMSE). More broadly, temporal models consistently achieved higher average correlations than the non-temporal models despite being trained on approximately an order of magnitude fewer samples, indicating a stronger ability to preserve relative Rmax variability across storms. This advantage becomes more pronounced when R34 is unavailable, suggesting that temporal information can partially compensate for missing storm-size predictors.

The study also investigates transfer learning using synthetic datasets (RAFT and STORM) for pre-training and IBTrACS data for fine-tuning. However, transfer learning does not improve performance, likely because the synthetic datasets contain lower and less variable Rmax distributions than the observational IBTrACS data. Distributional analysis revealed that the synthetic datasets are generally centered at lower Rmax values and exhibit a narrower spread than IBTrACS, meaning pre-trained models learned relationships that were not fully representative of observed storm behavior.

The paper concludes that "temporal deep learning models can effectively impute missing Rmax values, especially when physics-informed input variables are used, but there are trade-offs that arise due to the relatively smaller size of datasets available for training. The findings highlight both the promise of deep learning for improving historical storm records and the importance of matching synthetic training data to real observational distributions." Future work should investigate temporal architectures that dynamically incorporate available observed Rmax values, distribution-alignment or domain-adaptation methods for synthetic data, and evaluation frameworks that separate hyperparameter selection from final model assessment.

Improvements to AI Systems:

1. Physics-Augmented Temporal Imputation for Sparse Environmental Data

• Improvement: Integrate domain-specific physical parameters (e.g., R342, Coriolis parameter, sin(lat), Mmax/M34) as auxiliary inputs into LSTM/1DCNN architectures.

• What the improved system can do: Accurately reconstruct missing storm-structure variables (e.g., Rmax) from partial best-track records, even when key predictors like R34 are absent, by leveraging temporal dependencies and physical constraints.

2. Temporal Context-Aware Missing-Value Recovery

• Improvement: Replace static imputation models with sequence-based models (LSTM/1DCNN) that learn storm evolution patterns over time.

• What the improved system can do: Preserve relative variability of imputed values across storms, producing more realistic and physically consistent time series for historical cyclone records, outperforming non-temporal models (XGBoost, ANN) on correlation and RMSE.

3. Synthetic-to-Real Domain Alignment for Pre-Training

• Improvement: Add distribution-matching layers (e.g., adversarial domain adaptation, quantile matching, or reweighting) to align synthetic training data (RAFT/STORM) with observational distributions (IBTrACS) before fine-tuning.

• What the improved system can do: Avoid performance degradation from distribution shift, enabling effective transfer learning when observational data are scarce, and generating imputations that match real-world Rmax ranges and variability.

4. Dynamic Incorporation of Observed Values in Sequence Models

• Improvement: Modify LSTM/1DCNN architectures to accept variable-length input sequences with missingness masks, allowing the model to condition on any available observed Rmax values during imputation.

• What the improved system can do: Impute missing Rmax values more accurately by exploiting partially observed storm histories, rather than treating each storm as fully missing or fully observed.

5. Uncertainty-Aware Imputation with Physics-Based Priors

• Improvement: Add a physics-based parametric model (e.g., CK22) as a prior or regularizer in the loss function, and output predictive distributions (e.g., via Monte Carlo dropout or ensemble LSTM).

• What the improved system can do: Provide confidence intervals for imputed Rmax values, flag low-confidence imputations for manual review, and ensure outputs remain within physically plausible bounds.

6. Robust Evaluation Framework for Model Selection

• Improvement: Separate hyperparameter tuning from final performance assessment using nested cross-validation or a holdout set that is never used for model selection.

• What the improved system can do: Produce unbiased performance estimates and avoid overfitting to specific test splits, leading to more reliable deployment in operational hurricane reanalysis.