Abstract
Severe typhoons, as extreme weather events, can cause a large number of casualties and property damage in coastal areas. There are mainly three kinds of methods for the prediction of severe typhoon formation: numerical-based, statistical-based, and machine learning-based methods. However, existing methods often fail to account for the imbalance between ordinary and severe typhoon samples, leading to significantly lower accuracy for the latter. This paper proposes the Unbalanced Severe Typhoon formation Prediction (USFP) framework based on transfer learning. The framework consists of a pre-learning model trained on a balanced dataset to acquire prior knowledge, followed by a re-learning model that applies this knowledge to real-world unbalanced data. By fusing 3D atmospheric, 2D sea surface, and 3D ocean hydrographic variables, the USFP framework demonstrates superior performance over the ECMWF's IFS model and existing machine learning methods across multiple oceanic regions.
Introduction: The Challenge of Class Imbalance
Predicting when a tropical cyclone will escalate into a "severe" typhoon is a high-stakes challenge for coastal safety and economic planning. While machine learning has made significant strides in weather forecasting, it faces a fundamental hurdle: severe typhoons are rare. In data science terms, this is a "class imbalance" problem. In the Western Pacific, for instance, ordinary typhoon samples outnumber severe ones by more than 10 to 1.
Traditional machine learning models trained on such raw data tend to be biased toward the majority class, leading to high false-negative rates for severe events. Furthermore, many existing models rely primarily on atmospheric data, ignoring the complex interaction between the sky and the sea. Deep-ocean variables, such as heat content and sub-surface currents, are critical for understanding storm intensification but are often underutilized.
Methodology: The USFP Framework
The authors propose a two-stage architecture designed specifically to handle data scarcity and class imbalance through transfer learning.
Stage 1: The Pre-learning Model
The team first constructed an artificial "balanced" dataset by sampling equal numbers of ordinary and severe typhoons. This stage teaches the model what a severe typhoon "looks like" without the distraction of a lopsided dataset. By training on this balanced set, the model extracts robust features that define severe intensification.
Stage 2: The Re-learning Model
The "prior knowledge" (the trained model architecture and weights) from the pre-learning model is then transferred to a second stage. This model is trained on the real-world, unbalanced dataset but utilizes an LSTM (Long Short-Term Memory) network to account for the temporal evolution of the storm. To further combat imbalance, the researchers designed a customized "UST" loss function that assigns higher penalties for misclassifying severe typhoons.
Multi-Modal Fusion and Architecture
The USFP framework integrates three distinct data types to provide a holistic view of the storm environment:
- 3D Atmospheric Data: Temperature, humidity, geopotential height, and wind components.
- 2D Sea Surface Data: Sea surface temperature (SST).
- 3D Ocean Hydrographic Data: Seawater temperature and velocity at various depths.
As illustrated in the architecture diagrams, 3D Convolutional Neural Networks (3DCNNs) are used to extract features from the volumetric atmospheric and ocean hydrographic data, while 2DCNNs are employed for sea surface variables. These features are then fused to create a comprehensive input for the final prediction layers.
Key Results and Insights
The USFP framework was evaluated against several deep learning baselines and the industry-standard IFS (Integrated Forecasting System) from ECMWF.
- Superior Accuracy: In the Western Pacific, the USFP achieved a ROC_AUC of 0.754, significantly higher than the numerical IFS model (0.546).
- Precision-Recall Gains: On the PR_AUC metric, which is more robust for imbalanced datasets, the USFP outperformed traditional ML models by a wide margin (e.g., 0.300 vs. 0.129 for ConvLSTM in the WP region).
- The Ocean as the Engine: The researchers found that adding ocean hydrographic variables consistently improved performance across all regions, confirming that deep-ocean heat content is a critical predictor for intensification.
Conclusion
The USFP framework demonstrates that transfer learning can effectively mitigate the "class imbalance" problem in severe weather prediction. By leveraging a "warm start" from balanced data and incorporating multi-modal oceanic variables, the model provides a significant improvement over traditional numerical models. Future work will explore the integration of satellite imagery and "physics-informed" neural networks to combine deep learning's predictive power with established meteorological laws.
Source Paper: Pan et al. (2023). USFP: An unbalanced severe typhoon formation prediction framework based on transfer learning. Frontiers in Marine Science. DOI: 10.3389/fmars.2022.1046964
Summary updated