WTC-HST: Wavelet Transform Convolutions and Hierarchical Spatial Transformers for Polarimetric SAR Image Classification

Polarimetric synthetic aperture radar (PolSAR) image classification is an important task in remote sensing. However, due to its complex scattering mechanism and high-dimensional features, current methods face challenges in dealing with multiscale features and global contextual information. In this article, we propose a wavelet transform convolution and hierarchical spatial transformer (WTC-HST) network for PolSAR image classification to effectively combine local frequency feature extraction and multi-scale global context modeling. First, the WTC model guides the network to adaptively learn the high and low frequency features in PolSAR images by deeply fusing the wavelet transform with the convolutional neural network (CNN). By using the channel scaling technique, WTC can efficiently extract the low-frequency shape feature and high-frequency texture feature representations of the polarization coherence matrix. Second, the HST module hierarchically extracts global self-attention features of multi-scale targets through a refined chunking mechanism, and preserves image spatial information via the convolutional projection to significantly enhance global feature representation. Finally, experiments on three real PolSAR datasets show that WTC-HST significantly outperforms the existing state-of-the-art CNN and transformer-based methods, and the overall accuracies can reach 97.55%–99.62%.