Scholarlink

<p>CNN-transformer hybrid architectures, designed to capture both global context and local spatial details, show promising performance in the semantic segmentation of remote sensing images. However, existing models either neglect the multispectral data or fail to highlight the unique spectral features crucial for discriminative tasks. In particular, these models fail to fully utilize near-infrared (NIR) wavelengths, which provide critical information for detecting vegetation, crops, and plant diseases. In response to this limitation, this paper proposes HiFormer+, a novel hybrid architecture that leverages channel-wise feature recalibration through a squeeze-and-excitation block. Specifically, HiFormer+ contains two novel components, including the SE block applied at the first CNN level output and a skip connection in the decoder phase, where the squeeze-and-excitation block allows for better spectral attention. This modification enables the model to effectively capture the complementary information in multispectral images by emphasizing important spectral channels while suppressing less informative ones. Extensive experiments on three remote sensing image sets, namely the satellite-based DSTL image set and the UAV-based RIT-18 and WYR image sets, demonstrate the proposed architecture’s ability to exploit NIR information effectively. HiFormer+ significantly outperforms the baseline HiFormer and all other state-of-the-art semantic segmentation models using NIR images on all benchmarks, manifesting the potential for being a suitable model for multispectral images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

HiFormer+: CNN-Transformer Dual-Fused Model for Improving the Discriminational Power of Multispectral Remote Sensing Images

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Irem Ulku

摘要

CNN-transformer hybrid architectures, designed to capture both global context and local spatial details, show promising performance in the semantic segmentation of remote sensing images. However, existing models either neglect the multispectral data or fail to highlight the unique spectral features crucial for discriminative tasks. In particular, these models fail to fully utilize near-infrared (NIR) wavelengths, which provide critical information for detecting vegetation, crops, and plant diseases. In response to this limitation, this paper proposes HiFormer+, a novel hybrid architecture that leverages channel-wise feature recalibration through a squeeze-and-excitation block. Specifically, HiFormer+ contains two novel components, including the SE block applied at the first CNN level output and a skip connection in the decoder phase, where the squeeze-and-excitation block allows for better spectral attention. This modification enables the model to effectively capture the complementary information in multispectral images by emphasizing important spectral channels while suppressing less informative ones. Extensive experiments on three remote sensing image sets, namely the satellite-based DSTL image set and the UAV-based RIT-18 and WYR image sets, demonstrate the proposed architecture’s ability to exploit NIR information effectively. HiFormer+ significantly outperforms the baseline HiFormer and all other state-of-the-art semantic segmentation models using NIR images on all benchmarks, manifesting the potential for being a suitable model for multispectral images.