Enhancing tropical cyclone intensity forecasting with explainable deep learning integrating satellite observations...

Area of Science:

• Meteorology
• Climate Science
• Artificial Intelligence

Background:

• Tropical cyclone (TC) intensity forecasting is challenging due to complex interactions with environmental factors and climate change uncertainties.
• Accurate prediction of TC intensity changes, especially rapid intensification, is crucial for disaster preparedness.

Purpose of the Study:

• To develop and validate a hybrid deep learning model for forecasting tropical cyclone intensity.
• To assess the model's performance against existing methods, particularly for different lead times and rapid intensification events.

Main Methods:

• Proposed a hybrid convolutional neural network (hybrid-CNN) integrating satellite-derived spatial features and numerical prediction model outputs.
• Validated hybrid-CNN forecasts against best track data for tropical cyclone categories and phases.
• Compared hybrid-CNN performance with forecasts from the Korea Meteorological Administrator (KMA).

Main Results:

• Hybrid-CNN forecasts demonstrated significant improvements over KMA-based forecasts, with skill score enhancements of 22% (24h), 110% (48h), and 7% (72h).
• For rapid intensification cases, hybrid-CNN showed substantial improvements: 62% (24h), 87% (48h), and 50% (72h) over KMA forecasts.
• Explainable deep learning confirmed the hybrid-CNN's capability in predicting TC intensity.

Conclusions:

• The hybrid-CNN model offers a promising advancement in tropical cyclone intensity forecasting.
• Integrating satellite data with numerical model outputs via deep learning enhances prediction accuracy and reliability.
• This approach contributes valuable tools for improving tropical cyclone warnings and disaster management.