Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Class Prior-Free Positive-Unlabeled Learning with Taylor Variational Loss for Hyperspectral Remote Sensing Imagery

Hengwei Zhao, Xinyu Wang, Jingtao Li, Yanfei Zhong

TL;DR

This work tackles PU learning in hyperspectral remote sensing under no class-prior knowledge, addressing the tendency of unlabeled data to dominate training. It introduces T-HOneCls, a Taylor variational loss that down-weights the unlabeled-gradient impact while preserving the variational bound, coupled with a self-calibrated optimization loop (KL-Teacher) to stabilize training with limited labeled data. The approach demonstrates superior macro F1 accuracy on 7 datasets (21 tasks) for hyperspectral imagery and strong OA performance on CIFAR-10 and STL-10, outperforming both class-prior-based and prior-free PU methods. Practical contributions include a clear analysis of Taylor-order effects, an EMA-based consistency mechanism, and a global proportional random stratified sampler to enhance training stability and generalization across domains.

Abstract

Positive-unlabeled learning (PU learning) in hyperspectral remote sensing imagery (HSI) is aimed at learning a binary classifier from positive and unlabeled data, which has broad prospects in various earth vision applications. However, when PU learning meets limited labeled HSI, the unlabeled data may dominate the optimization process, which makes the neural networks overfit the unlabeled data. In this paper, a Taylor variational loss is proposed for HSI PU learning, which reduces the weight of the gradient of the unlabeled data by Taylor series expansion to enable the network to find a balance between overfitting and underfitting. In addition, the self-calibrated optimization strategy is designed to stabilize the training process. Experiments on 7 benchmark datasets (21 tasks in total) validate the effectiveness of the proposed method. Code is at: https://github.com/Hengwei-Zhao96/T-HOneCls.

Class Prior-Free Positive-Unlabeled Learning with Taylor Variational Loss for Hyperspectral Remote Sensing Imagery

TL;DR

This work tackles PU learning in hyperspectral remote sensing under no class-prior knowledge, addressing the tendency of unlabeled data to dominate training. It introduces T-HOneCls, a Taylor variational loss that down-weights the unlabeled-gradient impact while preserving the variational bound, coupled with a self-calibrated optimization loop (KL-Teacher) to stabilize training with limited labeled data. The approach demonstrates superior macro F1 accuracy on 7 datasets (21 tasks) for hyperspectral imagery and strong OA performance on CIFAR-10 and STL-10, outperforming both class-prior-based and prior-free PU methods. Practical contributions include a clear analysis of Taylor-order effects, an EMA-based consistency mechanism, and a global proportional random stratified sampler to enhance training stability and generalization across domains.

Abstract

Positive-unlabeled learning (PU learning) in hyperspectral remote sensing imagery (HSI) is aimed at learning a binary classifier from positive and unlabeled data, which has broad prospects in various earth vision applications. However, when PU learning meets limited labeled HSI, the unlabeled data may dominate the optimization process, which makes the neural networks overfit the unlabeled data. In this paper, a Taylor variational loss is proposed for HSI PU learning, which reduces the weight of the gradient of the unlabeled data by Taylor series expansion to enable the network to find a balance between overfitting and underfitting. In addition, the self-calibrated optimization strategy is designed to stabilize the training process. Experiments on 7 benchmark datasets (21 tasks in total) validate the effectiveness of the proposed method. Code is at: https://github.com/Hengwei-Zhao96/T-HOneCls.
Paper Structure (26 sections, 27 equations, 13 figures, 14 tables, 2 algorithms)

This paper contains 26 sections, 27 equations, 13 figures, 14 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Works
  3. Deep Learning Based Classification for HSI
  4. Positive and Unlabeled Learning
  5. Other Weakly Supervised Learning Methods
  6. Class Prior-Free PU Learning Framework with Taylor Variational Loss
  7. Taylor Variational Loss
  8. Preliminaries
  9. Theoretical Analysis of Variational Loss
  10. Taylor Series Expansion for Variational Loss
  11. Self-calibrated Optimization
  12. Experimental Results and Analysis
  13. Experimental Settings
  14. Datasets
  15. Training Details
  16. ...and 11 more sections

Figures (13)

  • Figure 1: T-HOneCls: A Taylor series expansion-based variational framework for HSI PU learning.
  • Figure 2: The curves of loss and F1-score of the variational classifier and T-HOneCls with different positive samples in the training stage (taking the cotton in the HongHu dataset as an example). The first row show the curves of the variational classifier, and the second row show the curves of the classifier proposed in this paper. The less positive class training data, the faster the variational model collapses.
  • Figure 3: The curves of loss and OA on CIFAR-10 and STL-10 datasets.
  • Figure 4: The F1-score curves (cotton in the HongHu dataset) for the different order of the Taylor series.
  • Figure 5: The F1-score curves (cowpea in the HanChuan dataset, o=5) for the different components of KL-Teacher.
  • ...and 8 more figures