Scholarlink

Abstract <p>Hyperspectral image (HSI) classification plays a pivotal role in remote sensing applications, providing precise identification of land cover types through detailed spectral analysis. The present study evaluates the effectiveness of deep autoencoder architectures across multiple hyperspectral datasets, including Pavia Centre, Pavia University, and Botswana. The autoencoder models were optimized and assessed using various optimizers – Adam, SGD, Adadelta, and Adagrad – over different epoch settings. Experimental results demonstrate exceptional classification accuracy, notably achieving up to 98.848% on Pavia Centre with the SGD optimizer, and 99.81% on the Botswana dataset with Adagrad, highlighting the method’s broad applicability and robustness. Across datasets, SGD exhibited the most stable performance among the tested optimizers. The outcomes affirm that autoencoder-based deep learning approaches efficiently encode intricate spectral–spatial characteristics, substantially advancing accuracy and reliability in hyperspectral image analysis.</p> Research highlights <p><UnorderedList Mark="Bullet"> <ItemContent> <p>Developed a hybrid HyperspectralAE model that integrates autoencoder-based reconstruction with supervised classification, achieving robust spectral–spatial feature learning.</p> </ItemContent> <ItemContent> <p>Achieved state-of-the-art accuracies across multiple hyperspectral datasets that outperform existing autoencoder-based approaches.</p> </ItemContent> <ItemContent> <p>Demonstrated that optimizer choice (SGD, Adagrad) plays a crucial role in model stability and performance, providing practical insights for hyperspectral deep learning applications.</p> </ItemContent> </UnorderedList></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

Spectral mastery: deep autoencoder driven classification across hyperspectral datasets

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Bhaskar Das,
  • Dhananjoy Bhakta,
  • Lalan Kumar

摘要

Abstract

Hyperspectral image (HSI) classification plays a pivotal role in remote sensing applications, providing precise identification of land cover types through detailed spectral analysis. The present study evaluates the effectiveness of deep autoencoder architectures across multiple hyperspectral datasets, including Pavia Centre, Pavia University, and Botswana. The autoencoder models were optimized and assessed using various optimizers – Adam, SGD, Adadelta, and Adagrad – over different epoch settings. Experimental results demonstrate exceptional classification accuracy, notably achieving up to 98.848% on Pavia Centre with the SGD optimizer, and 99.81% on the Botswana dataset with Adagrad, highlighting the method’s broad applicability and robustness. Across datasets, SGD exhibited the most stable performance among the tested optimizers. The outcomes affirm that autoencoder-based deep learning approaches efficiently encode intricate spectral–spatial characteristics, substantially advancing accuracy and reliability in hyperspectral image analysis.

Research highlights

Developed a hybrid HyperspectralAE model that integrates autoencoder-based reconstruction with supervised classification, achieving robust spectral–spatial feature learning.

Achieved state-of-the-art accuracies across multiple hyperspectral datasets that outperform existing autoencoder-based approaches.

Demonstrated that optimizer choice (SGD, Adagrad) plays a crucial role in model stability and performance, providing practical insights for hyperspectral deep learning applications.