PolSAR image classification using complex-valued multiscale attention vision transformer (CV-MsAtViT)

Abstract

This paper Introduces a novel method for Polarimetric Synthetic Aperture Radar (PolSAR) image classification using a Complex-Valued Multiscale Attention Vision Transformer (CV-MsAtViT). The model incorporates a complex-valued multiscale feature fusion mechanism, a complex-valued attention block, and a Complex-Valued Vision Transformer (CV-ViT) to effectively capture spatial and polarimetric features from PolSAR data. The multiscale fusion block enhances feature extraction, while the attention mechanism prioritizes critical features, and the CV-ViT processes data in the complex domain, preserving both amplitude and phase information. Experimental results on benchmark PolSAR datasets, including Flevoland, San Francisco, and Oberpfaffenhofen, show that CV-MsAtViT achieves superior classification accuracy, with an overall accuracy (OA) of 98.35% on the Flevoland dataset, outperforming state-of-the-art models like PolSARFormer. The model also demonstrates efficient computational performance, minimizing the number of parameters while preserving high accuracy. These results confirm that CV-MsAtViT effectively enhances the classification of PolSAR images by leveraging complex-valued data processing, offering a promising direction for future advancements in remote sensing and complex-valued deep learning.

The codes associated with this paper are publicly available at https://github.com/mqalkhatib/CV-MsAtViT.