PaperlayerView exact revision on arXiv

Source-linked AI summary

Hyperspectral Image Classification in the Presence of Noisy Labels

Junjun Jiang, Jiayi Ma, Zheng Wang, Chen Chen, Xianming Liu

arXiv:1809.04212v2cs.CV

TL;DR

Noisy labels are difficult to avoid in hyperspectral classification and can mislead training while reducing accuracy. RLPA uses spectral-spatial constraints with repeated random label propagation and label fusion to cleanse the noise. It improves classification over noisy-label baselines, with average OA, AA, and Kappa gains of 9.18%, 9.58%, and 0.1043.

  • Problem

    Noisy labels are difficult to avoid in hyperspectral data and can mislead classifier training while severely reducing classification performance.

  • Method

    RLPA constructs an SSPTM from spectral-spatial constraints and repeatedly propagates labels from randomly selected clean samples before fusing the results.

  • Results

    9.18%, 9.58%, and 0.1043 are the reported RLPA-over-NLA gains in average OA, AA, and Kappa, respectively.

  • Takeaways & Limitations

    RLPA provides a general noisy-label cleansing approach that can be applied to current classifiers and is especially effective when label noise is large.

Abstract

from arXiv · show

Label information plays an important role in supervised hyperspectral image classification problem. However, current classification methods all ignore an important and inevitable problem---labels may be corrupted and collecting clean labels for training samples is difficult, and often impractical. Therefore, how to learn from the database with noisy labels is a problem of great practical importance. In this paper, we study the influence of label noise on hyperspectral image classification, and develop a random label propagation algorithm (RLPA) to cleanse the label noise. The key idea of RLPA is to exploit knowledge (e.g., the superpixel based spectral-spatial constraints) from the observed hyperspectral images and apply it to the process of label propagation. Specifically, RLPA first constructs a spectral-spatial probability transfer matrix (SSPTM) that simultaneously considers the spectral similarity and superpixel based spatial information. It then randomly chooses some training samples as "clean" samples and sets the rest as unlabeled samples, and propagates the label information from the "clean" samples to the rest unlabeled samples with the SSPTM. By repeating the random assignment (of "clean" labeled samples and unlabeled samples) and propagation, we can obtain multiple labels for each training sample. Therefore, the final propagated label can be calculated by a majority vote algorithm. Experimental studies show that RLPA can reduce the level of noisy label and demonstrates the advantages of our proposed method over four major classifiers with a significant margin---the gains in terms of the average OA, AA, Kappa are impressive, e.g., 9.18%, 9.58%, and 0.1043. The Matlab source code is available at https://github.com/junjun-jiang/RLPA

I. INTRODUCTION

Hyperspectral classifiers face pervasive noisy labels that can mislead training and reduce accuracy, while relatively little work addresses this setting. RLPA uses spectral-spatial knowledge and random label propagation to cleanse noise before classification.

  • Noisy labels arise from limited information, complex land cover, unreliable automated or inexperienced labeling, and disagreement among multiple experts.
  • Hyperspectral classification under noisy labels remains understudied despite noisy labels severely decreasing land-cover classification accuracy.
  • RLPA exploits spectral-spatial constraints to guide noisy-label cleansing within a label-propagation framework.
  • The method constructs an SSPTM using spectral similarity and superpixel-based spatial information, then propagates labels from randomly selected clean samples to unlabeled samples.
  • RLPA is designed as a general preprocessing solution that can be applied to current classifiers and is especially effective when label noise is large.

II. PROBLEM FORMULATION

The paper formulates hyperspectral classification as predicting unseen-pixel labels from training pixels with noisy labels, while modeling label corruption through a class-transition probability.

  • A hyperspectral image contains hundreds of nearly contiguous spectral bands, and classification predicts unseen-pixel labels from labeled pixels.
  • The transition probability ρ_jk denotes P(Ỹ_ik = 1|Y_ij = 1) for distinct classes j and k.
  • The noisy-label problem uses training data X and a noisy label matrix Ỹ to predict the label of an unseen pixel x_t.

III. INFLUENCE OF LABEL NOISE ON HYPERSPECTRAL IMAGE CLASSIFICATION

Increasing label noise progressively degrades classifier performance, with sensitivity varying by classifier and database. The results also indicate substantial value in cleansing noisy labels before classification.

  • As label noise increases from 0 to 0.9, the performance of all evaluated classifiers gradually declines.The study reports average OA over ten runs for NN, SVM, RF, and ELM.
  • RF and ELM are relatively robust to label noise, whereas NN and SVM are more sensitive because they rely on nearest samples and support vectors.
  • With equal training-sample counts, University of Pavia’s OA declines faster than Salinas Scene’s, where the class counts are C = 9 and C = 16, respectively.
  • When ρ ≤ 0.7, samples with true labels can often remain dominant, allowing a good classifier to retain satisfactory performance.
  • Comparisons with an ideal noiseless subset show considerable room for label-cleansing improvement and underscore the importance of preprocessing.
  • At ρ = 0.5 and C = 16, the ratio of correctly labeled to samples labeled as another class is 15:1.

IV. PROPOSED METHOD

The proposed label-cleansing method combines hyperspectral image priors with random label propagation to improve noisy training labels. It constructs a spectral-spatial probability transfer matrix and repeatedly propagates labels from randomly selected “clean” samples.

  • Overview of the Framework: The method extracts prior knowledge from the training set to regularize label cleansing before subsequent classification.The framework is designed to improve label quality rather than replace the downstream classifier.
  • Overview of the Framework: RLPA randomly removes labels from selected samples and predicts them through label propagation using the SSPTM.Samples are split into labeled and unlabeled subsets during each propagation round.
  • Overview of the Framework: The repeated random splitting and propagation assumes correctly labeled samples form the majority, allowing clean information to spread across the training set.The method relies on majority-clean labels; if incorrect labels dominate, clean labels cannot be reliably recovered.
  • Overview of the Framework: The SSPTM combines spectral similarity with superpixel-based spatial constraints so same-class samples propagate more readily than different-class samples.This design addresses the difficulty of constructing a useful transfer matrix from noisy labels alone.
  • Overview of the Framework: The spatial component uses entropy rate superpixel segmentation, while the number of superpixels is adaptively determined from image texture complexity.The method first applies PCA and uses texture information to guide the superpixel count.

2) Construction of Spectral-Spatial Regularized Probabilistic Transition Matrix:

The spectral-spatial regularized probabilistic transition matrix defines label propagation on an affinity graph constrained by homogeneous superpixel regions. Spectral similarity supplies within-region edge weights, while cross-region connections are suppressed.

  • Construction of Spectral-Spatial Regularized Probabilistic Transition Matrix: The affinity graph connects pixels within the same homogeneous region and assigns zero edge weight to pixels from different regions.This superpixel constraint limits propagation across region boundaries.
  • Construction of Spectral-Spatial Regularized Probabilistic Transition Matrix: Within a homogeneous region, edge weights are determined by the spectral similarity between pixels.The method uses Euclidean distance to measure spectral similarity.
  • Construction of Spectral-Spatial Regularized Probabilistic Transition Matrix: The variance parameter in the similarity measure is calculated adaptively from the mean variance of pixels within each homogeneous region.The region-adaptive variance adjusts the similarity calculation to local spectral variation.
  • Construction of Spectral-Spatial Regularized Probabilistic Transition Matrix: The transition matrix converts affinity-graph weights into propagation probabilities, with T_ij representing the probability of jumping from node j to node i.Larger edge weights make travel between nodes easier.

C. Random Label Propagation through Spectral-Spatial Neighborhoods

RLPA cleans noisy training labels by repeatedly propagating labels from randomly selected “clean” subsets over a superpixel-based spectral-spatial graph, then combining the resulting assignments by majority vote.

  • Random label propagation: RLPA randomly divides training samples into labeled and unlabeled subsets, propagates labels through a superpixel-based spectral-spatial affinity graph, and outputs cleaned labels.The labeled subset size is l = round(N ∗η), where η is the selected “clean” sample proportion.
  • Random label propagation: Each propagation step combines current label information with information received from neighboring pixels, controlled by α.The unlabeled samples initially have zero label assignments.
  • Experimental outcome: 83.11% average OA is reported for RLPA, compared with 73.93% for NLA, 73.20% for Bagging, and 77.09% for iForest.These averages cover three databases and four classifiers.
  • Random label propagation: Repeating random subset assignment and propagation produces multiple label assignments for every training sample.The procedure generates S label matrices and S labels for each sample.
  • Random label propagation: Majority vote combines the repeated propagated labels into the final cleaned label for each training sample.MVA returns the majority of a sequence of elements.
  • Experimental outcome: The number of noisy-label samples decreases as RLPA iterations increase under the three examined noise levels.The paper attributes this expected improvement to incorporating spatial information during propagation.

V. EXPERIMENTS

The experiments evaluate RLPA using three hyperspectral databases, compare it with four other methods, assess the SSPTM, and study parameter effects.

  • Experimental design: The experimental setup uses three hyperspectral image databases and compares the proposed results with four other methods.The section also describes an assessment of the SSPTM and parameter settings.
  • Experimental design: Table I reports the numbers of samples in the Indian Pines, University of Pavia, and Salinas Scene images.The table uses background color to distinguish different classes.
  • Experimental design: Figure 7 presents RGB composite images and ground reference information for Indian Pines, University of Pavia, and Salinas Scene.The databases are shown as panels (a), (b), and (c), respectively.

A. Database

RLPA is evaluated on three publicly available hyperspectral databases using randomized training and testing samples, with label noise introduced at controlled levels.

  • Database: The experiments use three publicly available hyperspectral image databases.The databases include Indian Pines, University of Pavia, and Salinas Scene.
  • Database: Indian Pines contains 16 land-cover types, 200 usable bands after removing 20 low-SNR bands, and approximately 10,249 labeled pixels.The scene has 145×145 pixels and was acquired by AVIRIS.
  • Database: Training and testing samples are randomly selected from available ground-truth maps; Indian Pines uses 10% for training, while the other databases use 50 samples per class.The remaining samples are used for testing.
  • Database: The classification maps compare four methods across four classifiers on Indian Pines at noise levels ρ = 0.1 and ρ = 0.5.Rows represent methods and columns represent classifiers.
  • Database: The experiments restrict label-noise comparisons to ρ ≤0.5, assuming more than half of training labels are correct.Each training label flips to another label with probability ρ.

B. Result Comparison

RLPA is compared with direct noisy-label training, Bagging, and iForest across three databases and four classifiers. It generally performs best, particularly at higher noise levels, while its average gains over NLA are 9.18% in OA, 9.58% in AA, and 0.1043 in Kappa.

  • Methods: RLPA repeats random label assignment and propagation, then fuses labels across iterations for noise cleansing.The experiments compare this procedure with NLA, Bagging, and iForest using NN, SVM, RF, and ELM classifiers.
  • Comparison methods: Bagging cannot improve performance over NLA and may perform worse because resampling reduces the available training samples.This behavior indicates that resampling alone does not address noisy labels effectively in these experiments.
  • Comparison methods: iForest is classifier- and database-dependent, performs well with NN, and can gain more than three percentage points over NLA on average.Because iForest detects noisy samples without cleansing their labels, its performance may still fall below NLA or Bagging.
  • RLPA results: RLPA outperforms comparison methods in almost all situations, especially at large noise levels, with gains over NLA reaching 10% for NN and SVM.The gain decreases to 3% for RF and ELM, showing classifier-dependent improvements.
  • Visual results: RLPA with ELM achieves the best visual classification performance among the compared methods, but its maps can show salt-and-pepper effects in smooth regions.The paper attributes this artifact to RLPA’s pixel-wise operation and suggests incorporating spatial constraints to fuse classification results.

C. Effectiveness of SSPTM

The spectral-spatial probability transfer matrix improves label propagation over a spectral-only matrix. Its spatial constraint produces a sparse, highly diagonal structure that limits propagation across different classes.

  • Quantitative comparison: SS-PTM provides a considerable quantitative advantage over S-PTM in OA-based comparisons.The comparison uses statistical OA curves for methods and classifiers across the evaluated databases.
  • Matrix structure: SS-PTM is sparse and highly diagonalized, whereas S-PTM is dense and non-diagonal.The spatial structure is associated with propagation within the same class, while S-PTM contains edges between samples with different labels.

D. Parameter Analysis

RLPA performance depends on the clean-sample proportion η and the balance parameter α. The paper evaluates their effects on average OA across three databases with ELM under ρ = 0.3.

  • Parameter roles: η denotes the proportion of clean samples among training samples, while α balances current labels against labels received from neighbors.The study selects both parameters empirically through grid search.

VI. CONCLUSION AND FUTURE WORK

The paper addresses hyperspectral classification with noisy labels using RLPA, which incorporates spectral-spatial prior information to cleanse labels. Experiments on three public databases support its effectiveness, while the study leaves realistic structured noise for future work.

  • Conclusion: RLPA targets the practical problem of hyperspectral image classification when training labels are noisy.The paper describes noisy labels as pervasive and capable of misleading classifier training and reducing classification performance.
  • Conclusion: RLPA incorporates spectral-spatial prior information to guide label propagation for label cleansing.The reported experiments evaluate the approach on three public databases.
  • Future work: Real label noise may be sample-dependent, class-dependent, or adversarial, unlike the uniformly random noise used in this study.The paper identifies handling such realistic noise, including difficult edge or similar-pixel cases, as future work.
Loading 1809.04212v2…