DeepCyclone-RI: A Hybrid Multi-Modal Deep Learning Architecture for Enhanced Hurricane Rapid Intensification Prediction Using Geostationary Satellite Imagery

Rapid Intensification (RI) of tropical cyclones—defined as an increase in maximum sustained winds of at least 30 knots within 24 hours—presents a critical challenge to coastal safety. Traditional forecasting models often struggle to capture the subtle, real-time structural evolution of storm cores preceding these events. This study introduces DeepCyclone-RI, a novel hybrid deep learning architecture designed to fuse high-frequency visual data from GOES-16 satellites with thermodynamic wind parameters, effectively mimicking the multi-modal intuition of human meteorologists. Unlike standard computer vision approaches, this research engineered a cloud-native data pipeline capable of streaming and sanitizing 5 years of historical data (2019-2023) directly from AWS S3, comprising 1,501 unique storm samples. The model employs a custom dual-branch architecture: a Convolutional Neural Network (CNN) to extract spatiotemporal features from Infrared (IR) imagery and a Multi-Layer Perceptron (MLP) to process scalar wind data. To address the extreme rarity of RI events (~3% prevalence), a strategic oversampling pipeline was implemented, successfully overcoming the 'Accuracy Paradox' where baseline models achieved high accuracy but failed to detect any intensification. Results from the scaled dataset demonstrate that DeepCyclone-RI achieved 100% Recall (Zero False Negatives) on the validation set, identifying every major intensification event. Furthermore, Gradient-weighted Class Activation Mapping (Grad-CAM) analysis confirms that the model correctly focuses on the eyewall and Central Dense Overcast (CDO) regions, validating its physical interpretability. This research establishes a proof-of-concept for a 'fail-safe' screening tool that prioritizes public safety by ensuring no potential threat goes undetected. Video Presentation: A supplementary video explaining this work is available on YouTube at: https://www.youtube.com/watch?v=a-dS1jqx7UE Key Innovations & Findings: Hybrid Fusion Architecture: Mathematically concatenates spatiotemporal visual features (CNN) with physical wind data (MLP) to model storm dynamics. Solved the "Accuracy Paradox": Overcame the class imbalance of rare RI events (~3%) using strategic oversampling, shifting from a flatlined baseline to a highly sensitive model. Zero False Negatives: Achieved 100% Recall on the validation set, aligning with the operational philosophy that "Recall is Sanity" in disaster management. Cloud-Native Engineering: Utilized AWS Open Data and pyproj for automated geospatial coordinate transformation and data streaming without massive local storage. Explainable AI (XAI): Verified model trust using Grad-CAM, proving the AI focuses on meteorological structures (Eye Wall/CDO) rather than image artifacts. Technical Stack: Python, TensorFlow/Keras, AWS S3 (Open Data), Xarray, Pyproj, OpenCV.