[Paper Review] Prediction of typhoon tracks using a generative adversarial network with observational and meteorological data

Mario Rüttgers, Sang-Seung Lee+1 more|arXiv (Cornell University)|Dec 5, 2018

Tropical and Extratropical Cyclones ResearchEarth and Planetary Sciences14 references3 citations

TL;DR

This study proposes a conditional generative adversarial network (cGAN) that fuses satellite imagery and reanalysis meteorological data to predict typhoon tracks 6 hours in advance. By incorporating physically meaningful variables like sea surface temperature and surface pressure, the model reduces average center prediction error by 29.7% to 67.2 km and improves image sharpness, enabling accurate forecasting of both typhoon center location and dynamic cloud structures.

ABSTRACT

Tracks of typhoons are predicted using a generative adversarial network (GAN) with observational data in form of satellite images and meteorological data from a reanalysis database. Time series of images of typhoons which occurred in the Korean Peninsula in the past are used to train the neural network. The trained GAN is employed to produce a 6-hour-advance track of a typhoon for which the GAN was not trained. The predicted image favorably identifies the future location of the typhoon center as well as the deformed cloud structures. The errors between predicted and real typhoon centers are measured quantitatively in kilometers. 65.5 % of all typhoon center predictions have an error of less than 80 km, 31.5 % lie within a range of 80 - 120 km and the remaining 3.0 % are above 120 km. The overall error is 67.2 km, compared to 95.6 km when only observational data are used as input. The cloud structure prediction is evaluated qualitatively. It is shown that the GAN is able to predict trends in cloud motion. It is found that adding physically meaningful meteorological data to satellite images improves the sharpness of predicted images.

Motivation & Objective

• To improve typhoon track prediction accuracy by integrating observational satellite data with physically meaningful meteorological reanalysis data.
• To reduce image blurriness in GAN-generated typhoon forecasts through inclusion of dynamic atmospheric variables.
• To evaluate the impact of multimodal input (satellite + reanalysis) on predicting both typhoon center movement and cloud motion dynamics.
• To assess the model’s performance on unseen typhoons affecting the Korean Peninsula, focusing on 6-hour advance predictions.
• To determine whether reanalysis data enhance the realism and temporal consistency of generated typhoon images.

Proposed method

• A conditional GAN (cGAN) is trained on chronologically ordered satellite images and reanalysis meteorological data from 66 typhoons affecting the Korean Peninsula.
• Input includes 4 sequential satellite images and corresponding reanalysis data (e.g., sea surface temperature, surface pressure, wind velocity) at 200 mb level.
• The generator network produces a 6-hour-ahead typhoon image, while the discriminator evaluates realism and spatial consistency of generated outputs.
• The loss function combines standard GAN loss with a perceptual loss (VGG-based) and a gradient loss (GDL) to reduce blurriness and improve structural fidelity.
• Model training uses data from 1993 onward, with 10 unseen typhoons used for testing to evaluate generalization.
• Meteorological data are sourced from the ECMWF ERA-Interim reanalysis database, providing physically consistent atmospheric conditions.

Experimental results

Research questions

• RQ1 Can combining satellite images with reanalysis meteorological data improve the accuracy of 6-hour typhoon center predictions compared to using satellite images alone?
• RQ2 To what extent does the inclusion of physical atmospheric variables reduce image blurriness in GAN-generated typhoon forecasts?
• RQ3 How well can the model predict dynamic features such as cloud rotation, eyewall formation, and jet stream effects in future typhoon images?
• RQ4 Does the model generalize to typhoons not included in the training data, particularly those with complex behaviors like sudden course changes or landfall?
• RQ5 How do different input configurations (e.g., number of satellite images, inclusion of reanalysis data) affect prediction performance?

Key findings

• The inclusion of reanalysis data reduced the average typhoon center prediction error from 95.6 km to 67.2 km, a 29.7% improvement.
• 65.5% of predicted typhoon center locations had an error of less than 80 km, compared to 42.4% in the prior model using only satellite images.
• Only 3.0% of predictions had an error exceeding 120 km, down from 25.5% in the previous model, indicating significantly improved reliability.
• The model successfully captured dynamic features such as the spinning motion of typhoons, cloud structure deformation, and the influence of jet streams in generated images.
• The addition of reanalysis data notably improved image sharpness and realism, especially in regions like the eyewall and outer cloud bands.
• The model demonstrated better generalization on unseen typhoons, particularly in predicting complex behaviors like landfall and rapid intensification.