Scholarlink

<p>The fusion of hyperspectral imaging (HSI) and Light Detection and Ranging (LiDAR) data improves land-cover classification by combining complementary spectral and geometric information. However, many existing fusion frameworks use complex architectures and elaborate cross-modal interaction mechanisms, leading to high computational costs and limited scalability in large-scale or real-time hyperspectral-LiDAR processing systems. In this work, we study HSI–LiDAR classification from an accuracy–efficiency trade-off perspective and investigate how lightweight design can maintain competitive performance under strict per-node computational budgets in large-scale and distributed hyperspectral-LiDAR processing. Specifically, we adopt depthwise separable convolution (DSConv) in a modality-aware way. For HSI data, pointwise convolution compresses redundant spectral channels. For LiDAR data, spatially selective DSConv enhances geometric and texture representations. To model global dependencies with lower complexity, we use a simplified cross-modal fusion strategy: token concatenation followed by a shared Transformer encoder. We evaluate our method on three benchmark datasets—Houston2013, MUUFL, and Trento—and achieve overall accuracies of 90.97%, 93.26%, and 98.32%, respectively, with a model size of only about 1.1&#xa0;MB. Compared with more complex fusion models, the proposed method achieves consistently lower floating-point operations (FLOPs) while maintaining competitive classification accuracy and a compact parameter size. Rather than pursuing absolute superiority in all efficiency metrics, the model is designed under constrained computational budgets to achieve a favorable balance between accuracy and resource consumption. By reducing per-patch computational cost and memory footprint, the framework becomes computationally favorable for large-scale tile-based processing pipelines and parallel deployment scenarios. These characteristics imply strong scalability potential for high-throughput hyperspectral–LiDAR processing systems, where millions of spatial patches may need to be processed efficiently under limited hardware resources.</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

Efficient classification network for hyperspectral and LiDAR fusion based on depthwise separable convolution

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Changhong Yu,
  • Shuting Chen,
  • Binxiang Yu

摘要

The fusion of hyperspectral imaging (HSI) and Light Detection and Ranging (LiDAR) data improves land-cover classification by combining complementary spectral and geometric information. However, many existing fusion frameworks use complex architectures and elaborate cross-modal interaction mechanisms, leading to high computational costs and limited scalability in large-scale or real-time hyperspectral-LiDAR processing systems. In this work, we study HSI–LiDAR classification from an accuracy–efficiency trade-off perspective and investigate how lightweight design can maintain competitive performance under strict per-node computational budgets in large-scale and distributed hyperspectral-LiDAR processing. Specifically, we adopt depthwise separable convolution (DSConv) in a modality-aware way. For HSI data, pointwise convolution compresses redundant spectral channels. For LiDAR data, spatially selective DSConv enhances geometric and texture representations. To model global dependencies with lower complexity, we use a simplified cross-modal fusion strategy: token concatenation followed by a shared Transformer encoder. We evaluate our method on three benchmark datasets—Houston2013, MUUFL, and Trento—and achieve overall accuracies of 90.97%, 93.26%, and 98.32%, respectively, with a model size of only about 1.1 MB. Compared with more complex fusion models, the proposed method achieves consistently lower floating-point operations (FLOPs) while maintaining competitive classification accuracy and a compact parameter size. Rather than pursuing absolute superiority in all efficiency metrics, the model is designed under constrained computational budgets to achieve a favorable balance between accuracy and resource consumption. By reducing per-patch computational cost and memory footprint, the framework becomes computationally favorable for large-scale tile-based processing pipelines and parallel deployment scenarios. These characteristics imply strong scalability potential for high-throughput hyperspectral–LiDAR processing systems, where millions of spatial patches may need to be processed efficiently under limited hardware resources.