DOI: https://doi.org/10.36347/sjet.2026.v14i05.001

Pages: 189-193

The current remote-sensing satellites and ground-based systems of observations generate meteorological images at a high spatial resolution and temporal frequency, and manual interpretation becomes less and less acceptable in real-time. Existing deep learning algorithms are too expensive in terms of computation and slow to be trained on such data. In order to address this, a vision transformer-based (WCViT-LL) architecture is created using low-frequency wavelet compression. Haar wavelet transform divides the input into low and high-frequency subbands, and the LL component (where most of the semantic information is found) is retained, which decreases the data volume and sequence length at the input stage. This tiny representation is then fed through an ordinary Vision Transformer (ViT) in order to obtain features, which can be utilized to categorize the meteorological images. Complexity analysis shows that the calculation of WCViT-LL of the self-attention is about 6.25% of that of a typical ViT, which leads to faster inference. This paper presents a solution to the real-time processing and edge deployment of high-resolution meteorological images in a concise and uncomplicated way.