Scholarlink

<p>Hyperspectral Imaging (HSI) is an advanced imaging technique, and Deep Learning (DL) networks have been widely used for classifying them. Transfer learning approaches rely on pre-trained networks developed from public imagery datasets, which predominantly capture spatial features while underutilizing spectral information unique to HSI. This study evaluates the extent of spatial knowledge transfer in pre-trained and application-specific CNN models across agricultural, urban and mixed landcover environments. Experimental results indicate that while pre-trained DL networks achieve reasonable classification accuracy, they primarily rely on spatial feature abstraction, limiting spectral discrimination. However, introducing just 5% supplication-specific training data significantly enhances spectral feature utilization, enabling better functional classification. The findings highlight the potential for integrating minimal spectral domain knowledge into transfer learning frameworks to improve HSI classification. This study contributes to developing more efficient, semi-automated and expert-free HSI analysis techniques by optimizing the balance between spatial and spectral feature learning in DL models.</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

Functional Dynamics of the Knowledge Transfer and Pre-Training in Deep Learning Approaches for Hyperspectral Image Classification

  • 期刊论文

  • 发表日期：

    2025年3月27日

查看全文
  • Vamshi Krishna Munipalle,
  • Usha Rani Nelakuditi,
  • Rama Rao Nidamanuri

摘要

Hyperspectral Imaging (HSI) is an advanced imaging technique, and Deep Learning (DL) networks have been widely used for classifying them. Transfer learning approaches rely on pre-trained networks developed from public imagery datasets, which predominantly capture spatial features while underutilizing spectral information unique to HSI. This study evaluates the extent of spatial knowledge transfer in pre-trained and application-specific CNN models across agricultural, urban and mixed landcover environments. Experimental results indicate that while pre-trained DL networks achieve reasonable classification accuracy, they primarily rely on spatial feature abstraction, limiting spectral discrimination. However, introducing just 5% supplication-specific training data significantly enhances spectral feature utilization, enabling better functional classification. The findings highlight the potential for integrating minimal spectral domain knowledge into transfer learning frameworks to improve HSI classification. This study contributes to developing more efficient, semi-automated and expert-free HSI analysis techniques by optimizing the balance between spatial and spectral feature learning in DL models.