IEEE Geoscience and Remote Sensing Magazine - September 2019 - 76

bare soil. The water class had a commission rate of 56.82%
and a user accuracy of 43.18%. The results of the five classification methods are shown in Table 1, which presents the
overall accuracy, average accuracy, and Kappa coefficient.
OBJECT-BASED METHODS
Object-based techniques offer more advantages than pixelbased ones for UIS detection. Object-based methods can
apply multiple features that include semantic, geometric,
textural, topological, and spectral information. Multifeature combinations can reduce impulse noise and confusion
derived from the same object with different spectra and
different objects with the same spectra [153], [154]. The
minimum processing unit of object-based methods is an
object embedded with semantic information from an adjacent object aggregation rather than a single pixel. Objectbased procedures have been confirmed as efficient for UIS
extraction from high-spatial-resolution lidar data [155].
Lidar data are represented by point clouds recording multiechoes and intensities with accurate surface elevation information and building outlines. These significant geometric
and spatial features are effectively extracted by object-based
approaches as the basis of ground object classification. A
rule-sets-based [156] method and a threshold-based [157]
approach were proposed in the object-based framework for
building extraction from lidar data. In the two techniques,
ground objects were taken as objects while the geometric
and boundary features were used to build a classification.
Object-based methods use rule-based fuzzy classification and multiscale image segmentation, which extract
spatial information, textural features, and spectral information for UIS detection from high-resolution images.
This approach typically involves the following three steps:
1) selecting objects from adjacent pixels and identifying
interesting spectral elements
2) image segmentation, classification, and recombination
3) feature extraction.
However, users should know how to select the segmentation scale and which features must be selected before using
an object-oriented method. These two factors substantially
determine classification accuracy and computational efficiency [158]. In local-scale UIS detection, attention should
be paid to shadows and shelters generated by tall vegetation
and buildings in high-spatial-resolution RSIs.

TABLE 1. A CLASSIFICATION EVALUATION OF DIFFERENT
METHODS FOR THE PURPLE MOUNTAINS IN NANJING, CHINA.
METHODS

EVALUATION
INDEX

SVM

ANNs

DTs

K-MEANS

OA (%)

97.36

97.75

89.38

96.88

80.93

AA (%)

94.74

95.45

90.15

94.76

66.85

Kappa

0.95

0.96

0.85

0.94

0.65

OA: overall accuracy; AA: average accuracy; Kappa: Kappa coefficient.

Moreover, multiple-method fusion has shown good
performance in high-resolution image classification [159],
[160]. A traditional single algorithm and multimethod
combination have been evaluated [161]. For instance, a
combination of object- and pixel-based approaches was
used to extract UIS in Hangzhou city, China, from QuickBird images. The fusion of techniques adopted a segmentation masking strategy to overcome image analysis transformation from the pixel level to the object level. Pixel-based
analysis was used to acquire prior knowledge, and objectbased analysis identified similar-feature pixels using prior
knowledge by the weighted minimum distance method.
The method fusion achieved an overall accuracy of 91.9%,
a producer accuracy of 93.8%, and a user accuracy of
90.3% [162].
DEEP-LEARNING METHODS
Deep-learning procedures are used to establish and
simulate neural networks for human-like learning and
analysis and to translate low-level features into more abstract, high-level attribute features [152]. Deep learning
refers to a multilayer perception network with multiple
hidden layers, which can establish sample training and
testing by unsupervised learning methods, offer useful
identification features, and improve classification accuracy. Typical algorithms include the autoencoder [163],
deep Boltzmann machines [164], convolutional neural
networks (CNNs) [165], [166], and DBNs [167], [168].
In these techniques, the segmentation and merging of
images results in a dendrogram of hierarchical clustering, with space complexity O(n 2) and time complexity
O(n 3). These approaches are generally adopted in regional and local UIS detection because of high computing costs. Convolutional DBNs (CDBNs) [169] use an
advanced deep-learning algorithm. However, relatively
little research on UIS has been based on this method.
Therefore, in the following, we try to use an improved
CDBN method to extract UIS and obtain better results
compared with the SVM, CNN, and CDBN algorithms.
The CDBN method has a hierarchical convolution for
a 2D image structure based on DBNs. It is a hierarchical
generative model with a multilayer architecture for spatial
feature extraction and multilayer edge detection. This technique has three layers: an input layer of visible units v, a
detection layer of hidden units h, and a pooling layer of
probabilistic max-pooling p. A weight matrix W connects
the symmetrically visible layer and hidden layer. The visible
layer has an N V # N V matrix of binary units, representing
the data input for the algorithm. The detection and pooling layers each have K groups. Each group of hidden layers consists of an N H # N H matrix of binary units, and the
pooling layer has N P # N P binary units. Each K group has
N W # N W feature filter units, which connect the visible
and hidden layers. The probability of the two kinds of filter
units being calculated using multiple convolution operations is defined by the energy equation E(v, h) as follows:
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

SEPTEMBER 2019

IEEE Geoscience and Remote Sensing Magazine - September 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2019