CNN Profiler on Polar Coordinate Images for Tropical Cyclone Structure Analysis

Boyo Chen, Buo-Fu Chen, Chun-Min HsiaoAAAI 2021 · DOI

Introduction

While meteorologists have achieved high accuracy in predicting the tracks of tropical cyclones (TCs), understanding their internal wind structure remains a significant hurdle. Parameters such as maximum wind intensity ($V_{max}$) and the radius of 34-knot winds ($R_{34}$) are critical for numerical weather prediction models. Traditionally, these are estimated subjectively or via satellite-borne radar (ASCAT) with limited temporal coverage. This research introduces a novel CNN-based profiler that estimates the full radial wind structure of TCs using multi-channel satellite imagery, providing more accurate and physically consistent storm analysis.

Methods: The Polar Advantage

The core innovation of this approach is the shift from standard Cartesian $(x, y)$ image processing to a polar coordinate $(r, heta)$ framework.

Polar Coordinate Transformation

By projecting satellite images (Infrared, Water Vapor, Visible, and Passive Microwave) into polar coordinates, the spiral arms of a cyclone are transformed into roughly linear features. This alignment makes it significantly easier for the convolutional neural network (CNN) to learn structural patterns, as rotation in polar space becomes a simple linear shift (a "roll" operation).

Comparison of TC images in Cartesian vs. Polar coordinates across different spectral channels.

Consistent Multi-Task Learning

Existing automated methods often predict intensity and size as independent scalars. This can lead to physically contradictory results, such as a very small storm radius paired with an impossibly high wind speed. To solve this, the researchers developed a "profile-first" inference method. The CNN is trained to predict the entire radial wind profile. Specific parameters like $V_{max}$ and $R_{34}$ are then derived mathematically from this predicted profile, ensuring meteorological consistency.

Comparison of simultaneous multi-task learning (which can lead to contradictions) versus the proposed profile-based inference.

The TCSA Dataset

The study introduces the Dataset of Tropical Cyclone Structure Analysis (TCSA), a comprehensive resource released to the public to foster further research:

  • Scope: 76,835 images covering 1,407 TCs globally (2004–2018).
  • Channels: IR1, WV, VIS, and PMW.
  • Labels: Intensity, size, radius of maximum wind (RMW), and calculated 1D radial wind profiles.

Key Results and Insights

The CNN Profiler demonstrates high competitive performance against state-of-the-art models:

  • Intensity Estimation: Achieved a Root Mean Square Error (RMSE) of 10.43 knots for $V_{max}$.
  • Size Estimation: Achieved a Mean Absolute Error (MAE) of 50.58 km for $R_{34}$.
  • Structural Accuracy: Qualitative comparisons against ASCAT radar observations showed that the model accurately captures wind distribution in the inner core, an area where other satellite-based methods frequently underestimate speeds.

The research highlights that combining Infrared (cloud top data) with Passive Microwave (rainband data) provides the most robust structural profiling. By regularizing the model with meteorological logic—forcing it to output a continuous profile—the researchers have created a more reliable tool for real-time cyclone analysis.

Limitations and Future Work

Current limitations include a reliance on parametric models that assume storm symmetry, whereas real-world cyclones are often asymmetric due to wind shear. Additionally, only 46% of the dataset could be equipped with valid profile labels, as weaker storms often lack a defined $R_{34}$. Future work will focus on incorporating temporal data to capture storm evolution and expanding the model to account for quadrant-based size asymmetry.

Summary updated

BibTeX
@inproceedings{chen2021cnn,
  author = {Boyo Chen and Buo-Fu Chen and Chun-Min Hsiao},
  title = {{CNN Profiler on Polar Coordinate Images for Tropical Cyclone Structure Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i2.16183},
}
Figure