Scholarlink

<p>Labelled data scarcity remains a longstanding challenge in hyperspectral image analysis, primarily due to high spectral dimensionality and the laborious nature of manual annotation. Self-supervised contrastive learning (SSCL) recently emerged as a promising approach to address this challenge due to its ability to learn robust representations by distinguishing similar and dissimilar data samples guided by the inherent properties of data rather than by labels. Our study builds upon a previously established two-stage patch-level, multi-label classification method for hyperspectral imagery using contrastive learning and further examines its performance on single-label and multi-label classification tasks under scenarios of limited training data. The methodology unfolds in two stages. Initially, we train an encoder with a projection network in a contrastive learning approach. Next, we fine-tune the pre-trained encoder with a classifier. Our empirical results on four public hyperspectral datasets demonstrate consistent improvements over fully supervised methods, boosting overall accuracy by up to <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(4.4\%\)</EquationSource> </InlineEquation> in multi-label and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(7.14\%\)</EquationSource> </InlineEquation> in single-label tasks (e.g. on the Pavia University dataset). Performance remains competitive even under a <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(50\%\)</EquationSource> </InlineEquation> reduction in labelled training data. Our qualitative analysis confirms that the contrastive-based encoder can produce well-separated representations for different classes and identify location-based features, even though it was not explicitly trained on spatial cues. This suggests the method’s potential to uncover implicit spatial information. These findings highlight the value of self-supervised contrastive learning for hyperspectral image classification, offering a promising avenue for handling data scarcity while enhancing predictive accuracy. They also highlight the method’s resilience in real-world scenarios where abundant labelled examples are often unavailable.</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

Enhancing hyperspectral image prediction with contrastive learning in low-label regimes

  • 期刊论文

  • 发表日期：

    2026年2月4日

查看全文
  • Salma Haidar,
  • José Oramas

摘要

Labelled data scarcity remains a longstanding challenge in hyperspectral image analysis, primarily due to high spectral dimensionality and the laborious nature of manual annotation. Self-supervised contrastive learning (SSCL) recently emerged as a promising approach to address this challenge due to its ability to learn robust representations by distinguishing similar and dissimilar data samples guided by the inherent properties of data rather than by labels. Our study builds upon a previously established two-stage patch-level, multi-label classification method for hyperspectral imagery using contrastive learning and further examines its performance on single-label and multi-label classification tasks under scenarios of limited training data. The methodology unfolds in two stages. Initially, we train an encoder with a projection network in a contrastive learning approach. Next, we fine-tune the pre-trained encoder with a classifier. Our empirical results on four public hyperspectral datasets demonstrate consistent improvements over fully supervised methods, boosting overall accuracy by up to \(4.4\%\) in multi-label and \(7.14\%\) in single-label tasks (e.g. on the Pavia University dataset). Performance remains competitive even under a \(50\%\) reduction in labelled training data. Our qualitative analysis confirms that the contrastive-based encoder can produce well-separated representations for different classes and identify location-based features, even though it was not explicitly trained on spatial cues. This suggests the method’s potential to uncover implicit spatial information. These findings highlight the value of self-supervised contrastive learning for hyperspectral image classification, offering a promising avenue for handling data scarcity while enhancing predictive accuracy. They also highlight the method’s resilience in real-world scenarios where abundant labelled examples are often unavailable.