Neighbor-Consistent Transformer for Domain Adaptation in Remote Sensing Scene Classification

Unsupervised domain adaptation (UDA) has become a research hotspot in remote sensing scene classification (RSSC) to reduce the dependence on labeled samples and mitigate performance degradation caused by domain shift. Most existing UDA approaches rely on convolutional neural networks (CNNs), which are constrained by limited receptive fields and struggle to capture global contextual dependencies in complex remote sensing scenes. Vision transformers (ViTs) provide strong global modeling capability; however, their sequence-based representation lacks explicit constraints on spatial neighborhood structures, resulting in insufficient perception of fine-grained local structures, thereby causing attention misdirection under domain shift. To address this critical problem, a neighbor consistency modulation (NCM) module is first designed to extract local structural consistency representations and fuse them with global tokens, explicitly enhancing fine-grained land object perception. Furthermore, to achieve reliable category-level feature alignment, an adaptive probability pairwise alignment (APPA) method is introduced. In particular, an adaptive threshold is designed to retain reliable target pseudolabels, and discriminator-free pairwise constraints are proposed to align these retained samples with source samples via probability matching, thereby stabilizing cross-domain representations. Ultimately, these two complementary components are seamlessly integrated to form a novel transformer-based method, termed Neighbor-Consistent Transformer (NCT). Extensive experiments on three public remote sensing datasets demonstrate the superiority of the proposed NCT, achieving state-of-the-art performance in cross-domain adaptation. The code will be made publicly available at https://github.com/CccLYChen/NCT-UDA.