Modular Neural Networks with Fully Convolutional Networks for Typhoon-Induced Short-Term Rainfall Predictions

Area of Science:

• Meteorology and Atmospheric Science
• Artificial Intelligence and Machine Learning
• Geospatial Analysis

Background:

• Taiwan is highly susceptible to typhoons, experiencing significant natural disasters due to strong winds and heavy rainfall.
• Accurate prediction of typhoon-induced rainfall is crucial for disaster mitigation and management.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting hourly rainfall during typhoon events in Taiwan.
• To compare the performance of a novel model with conventional methods for typhoon rainfall forecasting.

Main Methods:

• Employed fully convolutional networks (FCNs), a deep learning technique for image recognition and semantic segmentation.
• Developed two FCN models: Ground Rainfall Image-based FCN (GRI_FCN) and a combined model using radar echo and ground rainfall data (GRI-RRI_MCNN).
• Utilized radar echo images and ground station rainfall data from southern Taiwan (2013-2019) and benchmarked against a Multilayer Perceptron (RMMLP).

Main Results:

• The GRI-RRI_MCNN model demonstrated a comprehensive understanding of future rainfall patterns during typhoons.
• This advanced model significantly enhanced the accuracy of hourly rainfall forecasting compared to the benchmark model.
• Evaluated forecast horizons ranging from 1 to 6 hours, confirming improved predictive capabilities.

Conclusions:

• The GRI-RRI_MCNN model offers a substantial advancement in predicting typhoon-related rainfall in Taiwan.
• This improved forecasting accuracy can lead to more effective disaster preparedness and response strategies.
• The study highlights the potential of deep learning, specifically FCNs, in meteorological forecasting for natural disaster management.