Hierarchical Bilinear Scattering Transformer for Remote-Sensing Image Scene Classification

Abstract

In remote-sensing scene classification (RSSC), persistent challenges such as high interclass similarity and substantial intraclass diversity remain key obstacles to accurate recognition. Although vision Transformer (ViT) has demonstrated outstanding performance, it tends to smooth out high-frequency discriminative details due to the inherent low-frequency preference of self-attention. To alleviate the feature oversmoothing induced by self-attention, we propose a novel hierarchical bilinear scattering Transformer (HBS-Former) that jointly captures local frequency-domain details and global semantic context. First, to alleviate the basis-dependent representation limitations of a single wavelet transform, a dual-stream scattering network is proposed by integrating the discrete wavelet transform (DWT) and the dual-tree complex wavelet transform (DTCWT). This design enables stable, direction-aware feature extraction in the frequency domain. Second, a global–local dual-path representation module is elaborately designed to capture long-range contextual information and fine-grained local details via global attention and local convolution branches, which effectively strengthens the complementarity of multigranularity feature representations. Furthermore, a hierarchical bilinear fusion (HBF) mechanism is developed to explicitly model second-order interactions among hierarchical features. By modeling pairwise bilinear interactions across hierarchical features, it produces more discriminative image representations. Extensive experiments on three public benchmarks, namely Aerial Image Dataset (AID), NWPU, and UCM, demonstrate that HBS-Former achieves a favorable tradeoff between classification accuracy and model complexity compared with existing RSSC methods. Code will be available at https://github.com/yizhilanmaodhh/HBS-Former