IEEE Geoscience and Remote Sensing Magazine - September 2019 - 85

efficiency; in this case, we recommend using other machine-learning algorithms. For example, the DT method is
not suitable for large-scale RSI processing because of high
computational costs [227]. The approach does not improve
the algorithm performance substantially for large-scale
data by optimizing parameters. By contrast, the RF, SVM,
and ANN methods have quicker training speeds compared
with DTs. Occasionally, the default parameter settings can
also fulfill the requirements for specific UIS detection.
Most machine-learning algorithms are supervised, although deep-learning algorithms can be either supervised or
unsupervised. For supervised machine-learning algorithms,
the quality and quantity of training samples directly affect
the learning and extraction accuracy. Collection of high-quality training samples should consider full-coverage samples
that have significant characteristics. The larger the number of
effective training samples, the better the learning effect, and
the higher the detection accuracy. However, in the real world,
we cannot obtain training samples of sufficient quality and
quantity because of various constraints. For instance, hyperspectral images contain fewer labeled data. Deep-learning
methods should be used in cases of fewer training samples or
uncertain quality. These approaches employ multilevel iterative learning, take the result of previous learning as the next
training sample, and self-optimize the training sample [171].
Compared with the SVM, DT, RF, and ANN methods,
deep-learning techniques have been shown to be more robust
and better suited to large or complex feature space data processing. Moreover, improving the predictive function of algorithms is helpful in solving the problem of less effective training samples, such as by creating a multiclassifier of machine
learning or combining regression analysis methods, both of
which may improve feature extraction accuracy. Additionally,
we can obtain many machine-learning software packages and
operating instructions free of charge from the Internet. Some
software packages support a variety of language environments
(e.g., MATLAB, R, and Python) and provide test data, which
are conducive to the learning and improvement of machinelearning algorithms to expand their range of applications.
AUTOMATIC DETECTION AND MULTIFEATURE
EXTRACTION
In the "Image Classification Methods" section, we referred
to the UIS automatic detection technique, which is useful in
practical applications, such as urban illegal building detection, ground object classification of RSIs, and UIS feature
extraction and mapping. UIS mappings can be established
automatically through the automatic detection system, and
new buildings can be obtained through comparison with
historical UIS mapping results while illegal buildings are
screened out [228]. Automatic UIS detection can effectively
decrease human interference factors and obtain noteworthy detection results, but the reliability and accuracy of the
algorithm should be guaranteed first.
Machine-learning methods, as mentioned previously,
have higher classification accuracy, strong adaptability and
SEPTEMBER 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

robustness, and great potential for automatic UIS detection.
In particular, deep-learning algorithms can extract multiple features from RSIs, even when the number of training
samples is small. In the "Image Classification Methods"
section, we proposed an improved CDBN method (dcCDBNs) and obtained anticipated experimental results for
UIS extraction. However, much remains to be done to refine this method for automatic UIS detection systems (e.g.,
nondestructive dimensionality reduction processing for hyperspectral images and fine boundary processing). Multifeature extraction and fusion may provide higher-accuracy
ground object identification and classification using deeplearning methods, providing a new potential feature extraction algorithm for automatic UIS detection.
CONCLUSIONS
UIS detection accuracy directly affects decision making in
urban management and the quality of environmental governance. To improve the accuracy of UIS detection, we must
be clear about the technical challenges and select effective
RSIs and methods. In this article, key studies related to UIS
detection were surveyed and summarized. We systematically
examined challenges in passive and active remote sensing,
proposed a general processing scheme, and outlined a classification strategy for the main algorithms and models in UIS
detection. Moreover, we verified and compared algorithm
performance, pointed out the advantages and disadvantages
of different types of algorithms, provided ideas for improving
the algorithms, and verified the performance of the proposed
improved deep-learning algorithm, dc-CDBNs. In addition,
we proposed a new framework regarding the relationships
among algorithms, application scenarios, detection scales,
and RSI spatial resolution for UIS detection. However, this
review does not examine all of the challenges and solutions
in detail, because new challenges constantly arise and corresponding methods are being implemented to address them.
ACKNOWLEDGMENTS
We would like to thank the anonymous reviewers for their
helpful comments and suggestions. We also thank Honglak
Lee for providing the CDBN source code and toolbox and
the U.S. Geological Survey for providing remote sensing
images. This work was financially supported by the National Natural Science Foundation of China under grant
31670552, and it was performed while the corresponding
author, Mingshi Li, acted as an awardee of the 2017 Qinglan Project, sponsored by Jiangsu Province, China.
AUTHOR INFORMATION
Yuliang Wang (ylw@chzu.edu.cn) received his M.S. degree in cartography and geographic information systems
from Xi'an University of Science and Technology, China, in
2009. He is a Ph.D. student in the College of Forestry, Nanjing Forestry University, China, and is also with the School
of Computer and Information Engineering, Chuzhou University, China. His research interests include urban remote
85

IEEE Geoscience and Remote Sensing Magazine - September 2019

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