Scholarlink

Despite their capability to capture precise characteristics, hyperspectral images (HSIs) have grown in popularity in a variety of applications due to the rich spectral and spatial information they provide. HSIs, as opposed to traditional images, provide a multidimensional representation, which improves discrimination and analytical abilities. This study explores how different activation functions affect the performance of deep learning (DL) models for hyperspectral image classification (HSIC), concentrating on metrics like overall accuracy ( \(\text {OA}\) ), actual accuracy ( \(\text {AA}\) ), and \(\text {Kappa}\) . The research highlights the importance of activation functions in DL model training for HSI processing, influencing convergence dynamics and generalization proficiency. HybridSN was chosen as the standard for assessing activation functions because of its competent integration of 2D and 3D convolutions for effective extraction of spectral and spatial data. Due to its basic architecture and consistent performance, HybridSN is well-suited for investigating the activation function’s effect on HSIC across varied datasets such as Indian Pines (IP), Pavia University (PU), and Salinas (SA). The paper compares the performance of ReLU, LeakyReLU, PReLU, ELU, SiLU, GELU, and Hardswish activation functions on classification tasks across several datasets, determining the most efficient activation functions for each dataset.

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

Finding Efficient Activation Functions for Deep Learning Based Hyperspectral Image Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Anish Sarkar,
  • Utpal Nandi,
  • Moirangthem Marjit Singh,
  • Bachchu Paul,
  • Rajasekaran Selvaraju

摘要

Despite their capability to capture precise characteristics, hyperspectral images (HSIs) have grown in popularity in a variety of applications due to the rich spectral and spatial information they provide. HSIs, as opposed to traditional images, provide a multidimensional representation, which improves discrimination and analytical abilities. This study explores how different activation functions affect the performance of deep learning (DL) models for hyperspectral image classification (HSIC), concentrating on metrics like overall accuracy ( \(\text {OA}\) ), actual accuracy ( \(\text {AA}\) ), and \(\text {Kappa}\) . The research highlights the importance of activation functions in DL model training for HSI processing, influencing convergence dynamics and generalization proficiency. HybridSN was chosen as the standard for assessing activation functions because of its competent integration of 2D and 3D convolutions for effective extraction of spectral and spatial data. Due to its basic architecture and consistent performance, HybridSN is well-suited for investigating the activation function’s effect on HSIC across varied datasets such as Indian Pines (IP), Pavia University (PU), and Salinas (SA). The paper compares the performance of ReLU, LeakyReLU, PReLU, ELU, SiLU, GELU, and Hardswish activation functions on classification tasks across several datasets, determining the most efficient activation functions for each dataset.