Scholarlink

<p>Recent advancements in remote sensing like satellites, airborne sensors, UAVs, etc. have enabled the data acquisition from multiple sources such as hyperspectral imaging (HSI) and LiDAR. Their effective fusion for land-use and land-cover (LULC) classification remains challenging due to heterogeneous feature spaces and distributional gaps. This research proposes DSSAGFNet, a unified deep learning framework, that integrates a dynamic spectral-spatial attention based HSI encoder and a graph based LiDAR encoder. The HSI encoder adaptively combines spectral and spatial attention features using learnable fusion weights, strengthened by residual and depthwise separable convolutions for efficient learning. While the LiDAR encoder employs a graph neural network leveraging EdgeConv and TopKPooling to capture detailed geometric features. To bridge the cross-modality differences, Maximum Mean Discrepancy (MMD) loss is applied, which improves feature alignment and fusion quality. Then transformer-inspired cross-modal attention mechanism facilitates the network to integrate complementary features, followed by a classification head to predict respective classes. Extensive experiments on Trento, Houston 2013 and MUUFL Gulfport datasets demonstrate that the proposed DSSAGFNet achieves competitive or superior performance compared to conventional classifiers, CNN-based methods and fusion based appraohes, exhibiting strong potential for real-world remote sensing applications.</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

DSSAGFNet: dynamic spectral-spatial attention and graph fusion network for hyperspectral-LiDAR data classification

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Ghulam Jilani Waqas,
  • Muhammad Kamran Saleem,
  • Danish Arif

摘要

Recent advancements in remote sensing like satellites, airborne sensors, UAVs, etc. have enabled the data acquisition from multiple sources such as hyperspectral imaging (HSI) and LiDAR. Their effective fusion for land-use and land-cover (LULC) classification remains challenging due to heterogeneous feature spaces and distributional gaps. This research proposes DSSAGFNet, a unified deep learning framework, that integrates a dynamic spectral-spatial attention based HSI encoder and a graph based LiDAR encoder. The HSI encoder adaptively combines spectral and spatial attention features using learnable fusion weights, strengthened by residual and depthwise separable convolutions for efficient learning. While the LiDAR encoder employs a graph neural network leveraging EdgeConv and TopKPooling to capture detailed geometric features. To bridge the cross-modality differences, Maximum Mean Discrepancy (MMD) loss is applied, which improves feature alignment and fusion quality. Then transformer-inspired cross-modal attention mechanism facilitates the network to integrate complementary features, followed by a classification head to predict respective classes. Extensive experiments on Trento, Houston 2013 and MUUFL Gulfport datasets demonstrate that the proposed DSSAGFNet achieves competitive or superior performance compared to conventional classifiers, CNN-based methods and fusion based appraohes, exhibiting strong potential for real-world remote sensing applications.