Satellite image-based Philippine Tropical Cyclone intensity classification using convolutional neural networks

Abstract

The Philippines experiences an average of 20 tropical cyclones (TC) each year, five of which are usually destructive, thus classifying TC intensity is crucial for disaster preparedness and protection of life and property. At present, the standard basis, Dvorak technique, uses grayscale images and heavily relies on the intuition of experts for TC feature extraction making it subjective with biases found to be a function of intensity. In this paper, we propose a convolutional neural network (CNN) model to automate TC intensity classification using color-enhanced images of TCs from 1985−2022 in the western North Pacific basin using satellite imagery. Initially, four architectures were tested using a raw color-enhanced dataset to assess model compatibility for the task. Transfer learning and fine-tuning experiments were incorporated to enhance the performance of the models. Out of all the architectures tested, the ResNet-50 model, with images preprocessed with fisheye distortion, which enhances TC key aspects, presented the best results based on validation accuracy.