PaperlayerView exact revision on arXiv

Source-linked AI summary

Spectral-Spatial Classification of Hyperspectral Image Using Autoencoders

Zhouhan Lin, Yushi Chen, Xing Zhao, Gang Wang

arXiv:1511.02916v1cs.CVcs.AIcs.LG

TL;DR

HSI classification commonly uses spectral information alone, despite the challenge of high-dimensional data with few labeled samples and the increasing importance of spatial information. This paper applies autoencoders to learn representations and proposes a spectral-spatial framework; its experiments report the highest accuracies for that framework.

  • Problem

    HSI classification often relies on spectral information alone, while high-dimensional features and limited labeled data challenge classification and spatial information is becoming more important.

    Source: §I · pp. 1–2

  • Method

    The paper uses single-layer and stacked autoencoders to learn representations and proposes a framework combining spectral and spatial information for classification.

    Source: §I · p. 2, §V · p. 8

  • Results

    The proposed spectral-spatial framework achieves the highest classification accuracies, and experiments report that deeper representations lead to better accuracies.

    Source: §V · pp. 8–9

  • Takeaways & Limitations

    The reported experiments support autoencoder-extracted representations as useful for HSI classification.

    Source: §V · p. 8

Abstract

from arXiv · show

Hyperspectral image (HSI) classification is a hot topic in the remote sensing community. This paper proposes a new framework of spectral-spatial feature extraction for HSI classification, in which for the first time the concept of deep learning is introduced. Specifically, the model of autoencoder is exploited in our framework to extract various kinds of features. First we verify the eligibility of autoencoder by following classical spectral information based classification and use autoencoders with different depth to classify hyperspectral image. Further in the proposed framework, we combine PCA on spectral dimension and autoencoder on the other two spatial dimensions to extract spectral-spatial information for classification. The experimental results show that this framework achieves the highest classification accuracy among all methods, and outperforms classical classifiers such as SVM and PCA-based SVM.

I. INTRODUCTION

HSI classification commonly relies on spectral information, but high-dimensional features and few labeled samples create a challenge, while spectral feature extraction has not consistently improved accuracy. This paper introduces autoencoder-based representations, including a framework that combines spectral and spatial information.

I. INTRODUCTION

  • High-dimensional features and few labeled samples make HSI classification vulnerable to the curse of dimensionality, and common feature-extraction methods have not consistently improved accuracy.In some cases, resulting accuracies fall below direct-classification approaches.Source: §I · p. 1
  • Most established methods use only pixel spectra, while spatial information is becoming more important as imaging-spectrometer resolution increases.Mathematical morphology is one approach cited for incorporating spatial information.Source: §I · p. 2
  • The paper introduces deep-learning-based feature extraction for HSI classification using single-layer and stacked autoencoders to learn shallow and deep representations.The representations are learned unsupervised and used for classification.Source: §I · p. 2
  • The proposed framework extracts spectral-spatial representations for HSI classification, and the methods are evaluated by classification accuracy.The paper also describes subsequent classification with logistic regression and SVM.Source: §I · p. 2

II. BACKGROUND

HSI classification assigns labels to pixels using their spectral information. Autoencoders learn representations by reconstructing inputs, and stacked autoencoders extend this process across multiple layers.

A. Hyperspectral Image Classification: Description

B. Autoencoders (AE) and Stacked autoencoders (SAE)

  • An autoencoder maps an input to a hidden latent representation, then maps that representation to an output matching the input size, called the reconstruction.The figure depicts the input x, hidden representation y, and reconstruction z.Source: B. Autoencoders (AE) and Stacked… · p. 3, Figure 2 · p. 3
  • A stacked autoencoder trains multiple autoencoder layers sequentially, with each subsequent layer trained on the previous layer’s output to produce a deep representation.The process is trained layer by layer.Source: B. Autoencoders (AE) and Stacked… · p. 3

III. LEARNING REPRESENTATIONS FOR HSI CLASSIFICATION

The paper extracts autoencoder representations and uses them as features for HSI classification, covering spectral-only and spectral-spatial approaches.

III. LEARNING REPRESENTATIONS FOR HSI CLASSIFICATION

  • The framework first computes autoencoder representations and uses them as data features before constructing a classifier for HSI classification.The section addresses both spectral-only schemes and a spectral-spatial framework.Source: §III · p. 3
  • For spectral-only classification, the paper proposes schemes using shallow and deep representations of HSI spectra.These schemes use different varieties of autoencoder-extracted representations.Source: §III · p. 3
  • The final scheme exploits both spectral and spatial information for classification.Source: §III · p. 3

A. Classifying with shallow spectral representation

The shallow spectral scheme uses an autoencoder to learn a representation, then trains an SVM on the hidden-layer activity. The autoencoder and SVM are trained separately.

A. Classifying with shallow spectral representation

B. Classifying with deep spectral representation

The paper uses stacked autoencoders to learn deep spectral representations, motivated by the variability of spectra within the same class. A logistic-regression classifier is placed on top of the learned representations.

B. Classifying with deep spectral representation

  • Within-class spectra vary with lighting, neighboring-object scattering, sensor rotation, and atmospheric conditions, motivating more robust and invariant representations.The paper argues that these variations make pixel-by-pixel analysis difficult.Source: B. Classifying with deep spectral… · p. 5
  • Higher-layer representations are proposed as more abstract and generally more invariant to local input changes.The paper presents stacked autoencoders as suitable for seeking such representations.Source: B. Classifying with deep spectral… · p. 5
  • Stacked autoencoders learn spectral representations, with logistic regression on top and fine-tuning across the architecture.The depicted six-layer instance has one input layer, four autoencoder hidden layers, and a logistic-regression output layer.Source: B. Classifying with deep spectral… · p. 5, Figure 3 · p. 5

C. Classifying with spectral-spatial representation

The spectral-spatial pipeline uses all pixels in a flat neighborhood rather than only a few tangent neighbors. It reduces spectral dimensionality with PCA before flattening the data for stacked autoencoders.

C. Classifying with spectral-spatial representation

  • The framework considers all pixels in a flat neighbor region, unlike approaches using only four or eight tangent neighbors and simple filtering.Autoencoders learn the representation from the neighborhood.Source: C. Classifying with spectral-spatial… · p. 5
  • A 7x7 neighborhood is extracted, but its hundreds of spectral channels create a very large, redundant input.The paper identifies this dimensionality as too large for the classifier.Source: C. Classifying with spectral-spatial… · p. 5
  • PCA retains the first three spectral components before flattening, reducing dimensionality while preserving spatial information.The transformation is fitted on the whole image; because PCA operates within pixels, some spectral information is discarded while spatial information remains intact.Source: C. Classifying with spectral-spatial… · p. 6, Figure 4 · p. 6
  • The flattened data enter a stacked autoencoder for layer-wise training, followed by fine-tuning with a final logistic-regression layer.The extraction sequence is PCA, flattening, then autoencoder representations.Source: C. Classifying with spectral-spatial… · p. 6, Figure 4 · p. 6

IV. EXPERIMENTAL RESULTS

Experiments evaluate the autoencoder-based classification method on two hyperspectral images from different environmental settings. The data are regularized before classification.

IV. EXPERIMENTAL RESULTS

  • The evaluation uses KSC, with 13 land-cover categories and 5,211 labeled pixels, and Pavia, with nine classes and 42,469 labeled pixels.KSC has 176 of its 224 spectral channels used; Pavia has 103 channels.Source: §IV · p. 6
  • All data are regularized to [0, 1] before classifier input.Experiments are conducted on an Ubuntu 12.10 system using an Intel i5-3230M processor; SVM and PCA use SciKit-Learn implementations.Source: §IV · p. 6

A. Single-layer spectral representation

On KSC and Pavia, AE-SVM outperforms direct SVM after enough training iterations, while performance is not sensitive to hidden-layer size. The paper also reports near-perfect reconstruction after several hundred epochs.

A. Single-layer spectral representation

B. Classifying with deep spectral representation

The spectral-spatial framework achieves the highest reported accuracies, while the deep spectral experiments find gains from depth and compare against traditional classifiers. The reported spectral-spatial error rates use three principal components from a 7×7 neighbor region.

B. Classifying with deep spectral representation

  • For KSC, the stacked network uses 176 spectral channels and hidden layers of size 100; for Pavia, it uses 103 channels and performs best with hidden layers of size 140.Source: B. Classifying with deep spectral… · p. 7
  • Deeper spectral representations reduce classification error, and the autoencoder methods show significant accuracy gains over SVM and PCA-based SVM comparisons.Table I’s hidden-layer count refers to the number of 100- or 140-sized layers.Source: Table I · p. 8, B. Classifying with deep spectral… · p. 8

C. Classifying with spectral-spatial representation

V. CONCLUSION AND DISCUSSION

The paper proposes a stacked-autoencoder framework that extracts both spectral and spatial information for hyperspectral image classification. Experiments report reduced SVM error and highest accuracies for spectral-spatial classification, with deeper representations also improving accuracy.

  • The proposed classification framework extracts spectral and spatial information using stacked autoencoders.Source: §V · p. 8
  • Autoencoder-extracted representations help lower SVM error, and deeper representations lead to better classification accuracy.The experiments also report that autoencoders are not sensitive to hidden-unit number.Source: §V · p. 8
  • The deep spectral-spatial framework achieves the highest reported accuracies in hyperspectral image classification.Source: §V · p. 9
  • Future work will incorporate other deep-learning models into the framework to further improve classification accuracy.Source: §V · p. 9
Loading 1511.02916v1…