Transformer-Based Cyclone Detection: Leveraging Spatiotemporal Features from Satellite Imagery

Tropical cyclone detection is critical for disaster preparedness and risk mitigation. This study introduces a transformer-based framework for identifying cyclones in meteorological satellite data, utilizing the Data-efficient Image Transformer (DeiT) architecture. Unlike traditional methods relying on temporally densified datasets, our approach operates on native sequential satellite observations, preserving cloud evolution dynamics. The primary observational method for tropical cyclone discovery and extensive atmospheric monitoring is satellite imaging. In contrast to many modern deep learning-based cyclone detection systems that depend on artificial temporal densification through optical-flow interpolation, frame synthesis, or constraint-based reconstruction, the suggested methodology only uses native sequential satellite data, maintaining the inherent temporal fidelity of cloud dynamics. For end-to-end cyclone classification, a transformer-based Data-efficient Image Transformer (DeiT) architecture is used, taking advantage of the long-range feature dependencies and global spatial context present in geostationary satellite data. In order to preserve historical cloud structure and motion continuity without adding artificial intermediate frames, the dataset, which consists of native North Indian Ocean satellite sequences, specifically incorporates previous temporal frames (t − 1, t − 2, etc.). Spatiotemporal dynamics are measured post-inference using physically interpretable descriptors that are only calculated between successive observations, as opposed to preprocessing-driven motion reconstruction. These include mean brightness variability, texture energy, and dense Farneback optical-flow magnitude, enabling objective characterization of cloud-top evolution and motion persistence. We integrate physically interpretable spatiotemporal descriptors, including mean brightness, texture energy, and dense optical flow magnitude, to characterize cloud-top evolution and motion persistence. Temporal differencing and lag-wise analysis support research-oriented visualization through line and scatter representations, facilitating examination of structural transitions and dynamical consistency across successive timesteps. Experiments on North Indian Ocean satellite sequences demonstrate competitive classification performance, with improved physical interpretability and reduced preprocessing complexity. Comparative evaluation against recent optical-flow-intensive cyclone detection studies indicates that the proposed framework achieves competitive classification performance while significantly reducing preprocessing complexity, eliminating interpolation-induced artifacts, and improving physical interpretability through explicit time-series characterization of cloud morphology and motion dynamics. Our framework offers a scalable solution for automated cyclone monitoring, enhancing early detection and continuous atmospheric surveillance. The implementation and code used in this study are publicly available at: https://doi.org/10.5281/zenodo.19249181