Scholarlink

While large quantities of image data are available, labels assigned to images are scarce, especially in land-cover/land-use classification of remotely-sensed images (RSI). Semi-supervised learning (SSL), which can exploit both labeled and unlabeled images, is thus becoming increasingly important in the classification of RSI. A recent approach successfully uses (ensembles of) predictive clustering trees (PCTs) for SSL in RSI classification, in conjunction with deep neural networks (DNNs) as feature extractors. The large number of features extracted by DNNs in this context raises serious computational complexity issues, as features are used as both inputs and targets by SSL PCTs. We address this problem by applying dimensionality reduction to the space of features extracted by the DNNs. When learning (ensembles) of PCTs for SSL classification, the reduced space is used as input instead of the original features. This greatly reduces the time complexity of SSL, while preserving the predictive performance. We verify that this holds for both PCTs and ensembles thereof, for different DNN architectures, on several multi-class classification datasets of RSI.

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

Dimensionality Reduction for Efficient Semi-supervised Learning from Remote Sensing Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sintija Stevanoska,
  • Marjan Stoimchev,
  • Jurica Levatić,
  • Sašo Džeroski

摘要

While large quantities of image data are available, labels assigned to images are scarce, especially in land-cover/land-use classification of remotely-sensed images (RSI). Semi-supervised learning (SSL), which can exploit both labeled and unlabeled images, is thus becoming increasingly important in the classification of RSI. A recent approach successfully uses (ensembles of) predictive clustering trees (PCTs) for SSL in RSI classification, in conjunction with deep neural networks (DNNs) as feature extractors. The large number of features extracted by DNNs in this context raises serious computational complexity issues, as features are used as both inputs and targets by SSL PCTs. We address this problem by applying dimensionality reduction to the space of features extracted by the DNNs. When learning (ensembles) of PCTs for SSL classification, the reduced space is used as input instead of the original features. This greatly reduces the time complexity of SSL, while preserving the predictive performance. We verify that this holds for both PCTs and ensembles thereof, for different DNN architectures, on several multi-class classification datasets of RSI.