[Paper Review] GlobeNet: Convolutional Neural Networks for Typhoon Eye Tracking from Remote Sensing Imagery

Seungkyun Hong, Seongchan Kim+2 more|arXiv (Cornell University)|Aug 11, 2017

Meteorological Phenomena and Simulations5 references26 citations

TL;DR

This paper proposes GlobeNet, a deep convolutional neural network (CNN) framework for end-to-end typhoon eye tracking from multi-channel infrared (IR) satellite imagery. By leveraging complex CNN architectures with inception modules and optimized activation functions (ELU/Tanh), the model achieves an RMSE of 0.02 in latitude-longitude prediction, equivalent to ~74.53 km in great circle distance, significantly outperforming simpler models.

ABSTRACT

Advances in remote sensing technologies have made it possible to use high-resolution visual data for weather observation and forecasting tasks. We propose the use of multi-layer neural networks for understanding complex atmospheric dynamics based on multichannel satellite images. The capability of our model was evaluated by using a linear regression task for single typhoon coordinates prediction. A specific combination of models and different activation policies enabled us to obtain an interesting prediction result in the northeastern hemisphere (ENH).

Motivation & Objective

• To develop a deep learning model that directly predicts typhoon eye coordinates from raw remote sensing satellite images without manual feature engineering.
• To evaluate the effectiveness of different CNN architectures and activation functions in capturing complex atmospheric dynamics from high-resolution IR imagery.
• To achieve accurate, real-time typhoon eye tracking using minimal computational overhead compared to traditional numerical weather prediction models.
• To explore the potential of end-to-end deep learning for visual weather event prediction using large-scale global satellite data.

Proposed method

• The model processes 4-channel infrared (IR) satellite images in NHWC format as input, with spatial resolution normalized to [0,1] across 80° latitude and 150° longitude.
• Two distinct CNN architectures are evaluated: a basic CNN and a complex CNN with multiple inception modules to extract hierarchical spatial features from cloud patterns.
• Each convolutional block applies ReLU, LeakyReLU, or ELU activation functions, followed by max-pooling to reduce spatial dimensions while preserving feature depth.
• The final feature maps are flattened and passed through three fully connected (dense) layers with nonlinear activation, followed by a single sigmoid-output layer for linear regression of typhoon center coordinates.
• The Adam optimizer with a learning rate of 1e-5 is used for training, and models are trained on a 9:1 training-to-testing split using Keras with TensorFlow backend.
• Model performance is evaluated using RMSE between predicted and ground-truth typhoon center coordinates (latitude and longitude), computed over 2,674 images from 2011–2016.

Experimental results

Research questions

• RQ1 Can a deep CNN effectively learn to localize the eye of a typhoon directly from high-resolution multi-spectral IR satellite imagery?
• RQ2 How do different activation functions (ReLU, LeakyReLU, ELU, Sigmoid, Tanh) impact the accuracy of typhoon center prediction in CNN-based models?
• RQ3 Does the inclusion of inception modules improve feature extraction and prediction performance compared to a basic CNN architecture?
• RQ4 Can end-to-end deep learning achieve competitive prediction accuracy with significantly reduced computational cost compared to traditional numerical weather prediction models?

Key findings

• The Complex CNN with ELU activation in convolutional layers and Tanh in dense layers achieved the best performance, yielding an RMSE of 0.02 in normalized coordinates.
• This RMSE corresponds to a prediction accuracy of approximately 74.53 km in great circle distance, representing a substantial improvement over baseline models.
• The Complex CNN consistently outperformed the Basic CNN across all activation function combinations, despite requiring roughly half the training samples due to higher memory consumption.
• The worst-performing model, Basic CNN with ReLU/Sigmoid, achieved an RMSE of 0.065, equivalent to ~362.91 km error, highlighting the importance of architectural and activation design.
• All models completed inference in a few seconds per sample, indicating suitability for real-time or near-real-time typhoon tracking applications.
• The results demonstrate that deep CNNs can effectively extract meaningful atmospheric features from raw satellite imagery for precise typhoon eye localization.