Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cross-View-Prediction: Exploring Contrastive Feature for Hyperspectral Image Classification

Anyu Zhang, Haotian Wu, Zeyu Cao

TL;DR

This work tackles the reliance on labeled data in hyperspectral image classification by introducing a self-supervised Cross-View-Prediction framework that constructs two semantic views via cross-channel prediction using VAE and AAE, followed by a BYOL-style contrastive learning stage. The method maximizes mutual information between views while minimizing conditional entropy, yielding compact, discriminative features that transfer well to downstream SVM classification. Across IP, PU, and SA datasets, the approach achieves state-of-the-art unsupervised performance, with parity split performing best among cross-representation strategies and contrastive learning providing further gains. The proposed framework offers a practical, annotation-light route to robust hyperspectral representations with clear potential for real-world remote sensing tasks.

Abstract

This paper presents a self-supervised feature learning method for hyperspectral image classification. Our method tries to construct two different views of the raw hyperspectral image through a cross-representation learning method. And then to learn semantically consistent representation over the created views by contrastive learning method. Specifically, four cross-channel-prediction based augmentation methods are naturally designed to utilize the high dimension characteristic of hyperspectral data for the view construction. And the better representative features are learned by maximizing mutual information and minimizing conditional entropy across different views from our contrastive network. This 'Cross-View-Predicton' style is straightforward and gets the state-of-the-art performance of unsupervised classification with a simple SVM classifier.

Cross-View-Prediction: Exploring Contrastive Feature for Hyperspectral Image Classification

TL;DR

This work tackles the reliance on labeled data in hyperspectral image classification by introducing a self-supervised Cross-View-Prediction framework that constructs two semantic views via cross-channel prediction using VAE and AAE, followed by a BYOL-style contrastive learning stage. The method maximizes mutual information between views while minimizing conditional entropy, yielding compact, discriminative features that transfer well to downstream SVM classification. Across IP, PU, and SA datasets, the approach achieves state-of-the-art unsupervised performance, with parity split performing best among cross-representation strategies and contrastive learning providing further gains. The proposed framework offers a practical, annotation-light route to robust hyperspectral representations with clear potential for real-world remote sensing tasks.

Abstract

This paper presents a self-supervised feature learning method for hyperspectral image classification. Our method tries to construct two different views of the raw hyperspectral image through a cross-representation learning method. And then to learn semantically consistent representation over the created views by contrastive learning method. Specifically, four cross-channel-prediction based augmentation methods are naturally designed to utilize the high dimension characteristic of hyperspectral data for the view construction. And the better representative features are learned by maximizing mutual information and minimizing conditional entropy across different views from our contrastive network. This 'Cross-View-Predicton' style is straightforward and gets the state-of-the-art performance of unsupervised classification with a simple SVM classifier.
Paper Structure (23 sections, 15 equations, 6 figures, 10 tables, 3 algorithms)

This paper contains 23 sections, 15 equations, 6 figures, 10 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related work
  3. Autoencoder based feature learning
  4. Unsupervised feature learning
  5. Proposed Method
  6. Cross representation learning
  7. VAE module
  8. AAE module
  9. Contrastive learning
  10. Maximize mutual information
  11. Minimize conditional entropy
  12. Update strategy
  13. Experiments and analysis
  14. Datasets
  15. Model implementation
  16. ...and 8 more sections

Figures (6)

  • Figure 1: Our cross-view-prediction strategy applies AAE and VAE to perform the cross channel prediction tasks to construct different views of original data for contrastive learning. Table in figure shows all of these methods achieve unsupervised state-of-the-art performance in IP dataset.
  • Figure 2: VAE structure for hypersepctral image
  • Figure 3: AAE structure for hypersepctral image
  • Figure 4: Structure of the system, sg means stop-gradient, $I(Z_1,Z_2)$ means mutual information between lantent $Z_1$ and $Z_2$, $H(X_2|X_1)$ means conditional entropy
  • Figure 5: t-SNE visualization of contrastive features in the IP dataset.
  • ...and 1 more figures