DeepLearning-BasedClassificationofHyperspectralData-深度学习文档类

需积分: 13 307 浏览量更新于2023-05-30 评论收藏 2.67MB PDF 举报

身份认证购VIP最低享 7 折!

领优惠券(最高得80元）

资源详情

资源评论

资源推荐

Deep Learning-Based Classiﬁcation

of Hyperspectral Data

Yushi Chen, Member, IEEE, Zhouhan Lin, Xing Zhao, Student Member, IEEE,

Gang Wang, Member, IEEE, and Yanfeng Gu, Member, IEEE

Abstract—Classiﬁcation is one of the most popular topics in

hyperspectral remote sensing. In the last two decades, a huge

number of methods were proposed to deal with the hyperspectral

data classiﬁcation problem. However, most of them do not hierar-

chically extract deep features. In this paper, the concept of deep

learning is introduced into hyperspectral data classiﬁcation for the

ﬁrst time. First, we verify the eligibility of stacked autoencoders by

following classical spectral information-based classiﬁcation.

Second, a new way of classifying with spatial-dominated informa-

tion is proposed. We then propose a novel deep learning framework

to merge the two features, from which we can get the highest

classiﬁcation accuracy. The framework is a hybrid of principle

component analysis (PCA), deep learning architecture, and logistic

regression. Speciﬁcally, as a deep learning architecture, stacked

autoencoders are aimed to get useful high-level features. Experi-

mental results with widely-used hyperspectral data indicate that

classiﬁers built in this deep learning-based framework provide

competitive performance. In addition, the proposed joint

spectral–spatial deep neural network opens a new window for

future research, showcasing the deep learning-based methods’ huge

potential for accurate hyperspectral data classiﬁcation.

Index Terms—Autoencoder (AE), deep learning, feature

extraction, hyperspectral data classiﬁcation, logistic regression,

stacked autoencoder (SAE), support vector machine (SVM).

I. INTRODUCTION

Y COMBINING imaging and spectroscopy technology,

hyperspectral remote sensing can get spatially and spec-

trally continuous data simultaneously. Hyperspectral data are

becoming a valuable tool for monitoring the Earth’s surface [1],

[2], and are used in a wide array of applications. An incomplete

list includes agriculture [3], mineralogy [4], surveillance [5],

physics [6], astronomy [7], chemical imaging [8], and environ-

mental sciences [9], [10]. A common technique in these applica-

tions is the classiﬁcation of each pixel in hyperspectral data. If

successfully exploited, the hypers pectral data can yield higher

classiﬁcation accuracies and more detailed class taxonomies

[11]. However, there are several critical problems in the classiﬁ-

cation of hyperspectral data: 1) curse of dimensionality, because

of the high number of spectral channels; 2) limited number of

labeled training samples; and 3) large spatial variabili ty of

spectral signature [12].

A lot of different classiﬁcation methods have been proposed

to deal with hyperspectral data classiﬁcation. Traditional

hyperspectral data classiﬁcation methods use spectral informa-

tion only, and the classi ﬁcation algorithms typically include

parallelepiped classiﬁcation, k-nearest-neighbors, maximum-

likelihood, minimum distance, and logistic regression [13]. The

majority of these above algorithms suffer a lot from the “curse

of dimensionality.” To deal with the high dimensionality

and limited training samples of hyperspectral data [14], some

dimensionality reduct ion-based classiﬁcation methods were pro-

posed. Transformation is one method available to deal with high

dimensionality [15]–[17]. Band selection is another method

available to mitigate this “curse” [18]–[20].

In [21], a promising classiﬁcation method, support vector

machine (SVM), is introduced for hyperspectral data classiﬁca-

tion. SVM exhibits low sensitivity to high dimensionality and is

unlikely to suffer from the Hughes phenomenon [22]. In most

cases, SVM-based classiﬁers can obtain better classiﬁcation

accuracy than other widely used pattern recognition techniques

[14], [22]. For a long time, these class iﬁers were the state-of-the-

art methods [23].

Spatial information has been growing more and more impor-

tant for hyperspectral data classiﬁcation in recent years [30].

Spatial–spectral classiﬁcation methods provide signiﬁcant ad-

vantages in terms of improving performance [10]. To deal with

spatial variability of spectral signature, some recent approaches

try to incorporate spatial information into consideration [31]–[33].

In [32], the proposed method based on the fusion of morphol-

ogical information and original data followed by SVM provides

good classiﬁcation results. In [33], a new classiﬁcation frame-

work is proposed to exploit the spatial and spectral infor-

mation using loopy belief propagation and active learning. In

recent years, sparse representation-based methods have been

widely used in m a ny ﬁelds. In [34], spatial–spectral kernel

sparse representation is proposed to deal with hyperspectral

data classiﬁcation.

Considering the machine learning task of classiﬁcation, clas-

siﬁers like linear SVM and logistic regression can be attributed to

single-layer classiﬁers, whereas decision tree or SVM with

kernels are believed to have two layers [24]. As is conﬁrmed

in neuroscience, human brains perform well in tasks like object

recognition because of its multiple stages of processing from

See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Manuscript received October 14, 2013; revised April 24, 2014; accepted May

30, 2014. Date of publication June 25, 2014; date of current version August 01,

2014. This work was supported in part by the Fundamental Research Funds for the

Central Universities under Grant HIT. NSRIF.2013028 and in part by National

Natural Science Foundation of China under Grant 61301206 and Grant

61371180. (Corresponding author: Yushi Chen.)

Y. Chen, Z. Lin, X. Zhao, and Y. Gu are with the Institute of Image and

Information Technology, Harbin Institute of Technology, Harbin 150001, China

(e-mail: chenyushi@hit.edu.cn; lin.zhouhan@gmail.com; zhaoxing@hit.edu.cn;

yfgu@hit.edu.cn).

G. Wang is with the School Electrics and Electronics Engineering, Nanyang

Technological University, 639798 Singapore (e-mail: wanggang@ntu.edu.sg).

Color versions of one or more of the ﬁgures in this paper are available online at

http://ieeexplore.ieee.org.

Digital Object Identiﬁer 10.1109/JSTARS.2014.2329330

2094 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 7, NO. 6, JUNE 2014

retina to cortex [25]. Similarly, machine learning systems with

multiple layers of processing extract more abstract, invariant

features of data, and thus are believed to have the ability of

yielding higher classiﬁcation accuracy than those traditional,

shallower classiﬁers. These deep architectures have been shown

to yield promising performance in many ﬁeld including classiﬁ-

cation or regression tasks that involve image [26], [27], [47],

language [28], and speech [29].

In this paper, we introduce deep learning-based feature

extraction for hyperspectral data classiﬁcation for the ﬁrst time.

Our work focuses on applying autoencoder (AE), which is one of

the deep architecture-based models, to learn deep features of

hyperspectral data in an unsupervised manner. Our methods

exploit single-lay er AE and multi-layer stacked AE (SAE) to

learn shallow and deep features of hyperspectral data, respec-

tively. Furthermore, we propose a new way of extracting spatial-

dominated information for classiﬁcation. At last, we propose a

novel classiﬁcation framework dealing with joint spectral–

spatial information, which utilizes all of the features extracted

in the former two sections.

The rest of this paper is organized into six sections. Section II

is a description of deep learning, AE, and SAE models used in

this pape r. In Section III, we focus on classifying with spectral

features, whereas Section IV details a new way of incorporating

spatial information by extracting spatial-dominated features. In

Section V, we further merge the former two spectral and spatia l

approaches and propose a novel joint spectral–spatial deep learn-

ing framework, which yields the highest classiﬁcation accuracy.

Experimental results are shown in Section VI. Section VII

summarizes the observ ations and completes this paper.

II. D

EEP LEARNING, AE, AND SAE

A. Deep Learning

As early as 1989, the universal expressive power of three-layer

nets was proved via bumps and Fourier ideas [35]. The proof

showed surprisingly that any continuous function from input to

output can be implemented in a three-layer net, given sufﬁcient

number of hidden units and proper nonlinearities in activation

function and weights. However, due to the lack of proper training

algorithms in early years, people could not harness this powerful

model until Hinton proposed his deep learning idea in 2006 [27].

Deep learning involves a class of models which try to hierar-

chically learn deep features of input data with very deep neural

networks, typically deeper than three layers. The network is ﬁrst

layer-wise initialized via unsupervised training and then tuned in

a supervised manner. In this scheme, high-level features can be

learned from low-level ones, whereas the proper features can be

formulated for pattern classiﬁcation in the end. Deep models can

potentially lead to progressively more abstract and complex

features at higher layers, and more abstract features are generally

invariant to most local changes of the input. According to some

recent papers [36], [37], deep models can give better approxi-

mation to nonlinear funct ions than shallow models.

Typical deep neural network architectures include deep belief

networks (DBNs) [38], deep Boltzm ann machines (DBMs)

[39], SAEs [40], and stacked denoising AEs (SDAEs) [41].

The layer-wise training models have a bunch of alternatives

such as restricted Boltzmann machines (RBMs) [42], pooling

units [43], convolutional neural networks (CNNs) [44], AEs,

and denoising AEs (DAE) [40]. In this paper, we adopt one of

the above deep learning models, AE, for hyperspectral data

classiﬁcation and choose SAEs as the corresponding deep

architecture.

B. Autoencoders

An AE has one visible layer of inputs, one hidden layer of

units, one reconstruction layer of d units, and an activation

function (Fig. 1).

During training, it ﬁrst maps the input

to the hidden layer

and produces the latent activity

. The network correspond-

ing to this step is shown in the boxed part of Fig. 1 and is called an

“encoder.” Then, is mapped by a “decoder” to an output layer

that has the same size of the input layer, which is called

“reconstruction.” The reconstructed values are denoted as

. Mathematically, these two steps can be formulated as

where

and denote the input-to-hidden and the hidden-to-

output weights, respectively,

and denote the bias of hidden

and output units, and denotes the activation function.

Conventionally, the nonlinearity is provided in . There are

a lot of alternatives for such as sigmoid function, hyperbolic

tangent, and rectiﬁed linear function.

In our paper, the following constraint holds

We say that the AE has tied weights, which helps to halve

model parameters. Thus, we have three groups of parameters

remaining to learn: ,

, .

The goal of training is to minimize the “error” between input

and reconstruction, i.e.,

where is dependent on parameters ,

, while is given.

stands for the “error,” which can be deﬁned in a variety of

Fig. 1. Single layer AE for hyperspectral data classiﬁcation. The model learns a

hidden feature “ ” from input “ ” by reconstructing it on “ .” Corresponding

parameters are denoted in the network.

CHEN et al.: DEEP LEARNING-BASED CLASSIFICATION OF HYPERSPECTRAL DATA 2095

ways. Thus, the weight updating rule can be deﬁned as (where

denotes the learning rate)

After training the network, the reconstruction layer together

with its parameters are removed and the learned feature lies in the

hidden layer, which can subsequently be used for classiﬁcation or

used as the input of a higher layer to produce a deeper feature.

The power of AE lies in this form of reconstruction-oriented

training. Note that during reconstruction, it only uses the infor-

mation in hidden layer activity , which is encoded as features

from input. If the model can recover original input perfectly from

, it means that retains enough information of the input. And the

learned nonlinear transformation, which is deﬁned by those

weights and biases, can be deemed as a good feature extraction

step. So, stacking the encoders trained in this manner minimizes

information loss. At the meantime, they preserve abstract and

invariant information in deeper feature. This is the reason why we

choose AE to progressively extract deep features for hyperspec-

tral data.

C. Stacked AE

Stacking the input and hidden layers of AEs together layer by

layer constructs a SAE. The model is used to generate deep

features of hyperspectral data. Fig. 2 shows a typical instance of a

SAE connected with a subsequent logi stic regression classi ﬁer.

The ﬁrst AE maps inputs in 0th layer to a ﬁrst layer feature in

ﬁrst layer. It is trained using the same method introduced in

Section II-B. After we ﬁnish training the ﬁrst layer AE, subse-

quent layers of AEs are trained via the output of its previous

layer. For example, although we are training the AE between the

second and third layer, we try to recons truct the output of the

second layer according to the activity of the third layer. After this

layer of training, the decoder of the third layer AE is cast away

and only the input-to-hidden parameters are incorporated as

weights between the second and the third layer.

If the subsequent classiﬁer is implemented as a neural network

too, parameters throughout the whole network can be adjusted

slightly while we are training the classiﬁer. This step is called

ﬁne-tuning. For logistic regression, the training is simply back

propagation, searching for a minimum in a peripheral region of

parameters initialized by the former step.

III. C

LASSIFYING WITH SPECTRAL FEATURES

There exist some motivations to extract robust deep spectral

features. First, because of the complex sit uation of lighting in the

large scene, objects of the same class show different spectral

characteristics in different locations. For example, a lawn ex-

posed to direct sunlight shows different spectral characteristics

from a similar lawn eclipsed from the sunlight by a high building.

Also, scattering from other peripheral ground objects tilts the

spectra of the lawn and changes its characteristics too. Other

factors involve rotations of the sensor, different atmospheric

scattering conditions, and so on. According to these factors, the

probability distribution of a certain class is hard to be one-hot and

has variations over multiple directions in the feature space. These

complex variations of spectra make it hopele ss to analyze pixel

by pixel how they are affected by their tangent pixels in the

complicated real situation, thus they demand more robust and

invariant features. It is believed that deep architectures can

potentially lead to progressively more abstract features at higher

layers of feature, and more abstract features are generally

invariant to most local changes of the input [24].

To get more generally invariant featu res and tackle these

problems, a deep spectral feature of hypers pectral data can be

learned progressively layer by layer with the aforementioned AE

models. Gene rally speaking, we ﬁrst compute features via a SAE

and deem them as the features of data, then construct a logistic

regression classiﬁer on top of the neural network to ﬁnish the

classiﬁcation phase. By adjusting different numbers of layers of

AEs, both shallow and deep features can be learned. Fig. 3 shows

a typical instance of the de ep architecture used in our paper. The

training procedure will be detailed below.

A. Hierarchal Pretraining

The ﬁrst stage is to learn a deep feature of spectra via

pretraining a SAE in a hierarchal manner, which is outlined in

Fig. 2. Instance of a SAE connected with a logistic regression layer. It has ﬁve

layers: one input layer, three hidden layers, and an output layer.

Fig. 3. Classifying with spectral feature. The classiﬁcation scheme shown here

has ﬁve layers: one input layer, three hidden layers of AEs, and an output layer of

logistic regression. If we want to learn a shallower feature set, we just remove the

higher layers of AE.

2096 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 7, NO. 6, JUNE 2014

剩余13页未读，继续阅读

monotonomo

粉丝: 13
资源: 11

会员权益专享

Deep Learning-Based Classification of Hyperspectral Data

评论0

会员权益专享

最新资源