Thunderstorm Pattern Analysis Using Vision Transformers and Optimal Transport for Aviation Operations

Thunderstorms pose significant challenges to aviation safety and operational efficiency, requiring advanced tools for accurate analysis and prediction. This research explores a novel framework integrating Vision Transformers (ViTs) and Optimal Transport (OT) theory to analyze large-scale, unlabelled thunderstorm datasets. Trained in an autoregressive manner, the ViT model extracts both spatial structures and temporal evolution across thunderstorm radar images, producing latent feature embeddings that represent thunderstorm dynamics. These embeddings are treated as empirical distributions (or data support in latent space), enabling the application of OT theory to quantify the similarity between thunderstorm sequences. The resulting similarity measures support comparative analysis and similar sequence retrieval. Experimental results demonstrate that this framework effectively identifies structurally and temporally similar thunderstorm events, with promising implications for enhanced pattern recognition, operational decision-making, and risk mitigation in aviation meteorology.