IEEE Geoscience and Remote Sensing Magazine - September 2019 - 83

TABLE 4. THE ADVANTAGES AND DISADVANTAGES OF VARIOUS METHODS IN ACCORDANCE WITH RSIs' SPATIAL RESOLUTION
(CONTINUED).
SPATIAL
DATA
RESOLUTION SOURCE
Multisource
data
fusion

MAJOR
METHODS

ADVANTAGES

DISADVANTAGES

REFERENCE

Spatial
based

Expert
systems-based
method, ISODATA/RF

Enhanced spatial resolution of the
whole object images; overcoming
lower spatial resolution of larger space
scale

Combining of multiple algorithms
leading to instability of image
features extraction and fusion

[191], [185]
[192]

Spatial-
spectral
based

RF/ principal
component
analysis

Spatial and spectral features extraction; multifeatures combined to
improve classification accuracy

Necessity of image features to be
predefined manually, leading to
increasing algorithm complexity

[186], [54]
[213]
[220]

Spatiotemporal
based

Unmixing based
Filter-based
method
Deep-learning
methods
(CNNs, CDBNs)

Multiscale and multilayer features
extraction by deep learning; merging
respective advantages of spatial, temporal, or spectral resolutions; generating higher accuracy classification

High computational complexity; high-performance computer
configuration

[197], [198]
[199]
[200]

CATEGORY

coarse-, medium-, and high-resolution data, all of which have
been applied in UIS detection under suitable classification
methods. Generally, SMA, linear SMA, and subpixel-based
techniques are suitable for medium- to low-resolution images in UIS detection. For medium- and high-resolution images, pixel-based (machine-learning) or object-based methods
are usually applied. Urban Index approaches are frequently
used in regional- and global-scale UIS detection from medium- to low-resolution RSIs.
Multisource data fusion models have been developed to
capitalize on the respective advantages of RSIs and related
fusion methods. These fusion models reduce confusion
from spectral heterogeneity in landscapes and enhance the
classification accuracy in multifeature extraction. Regarding
image resolution, these models can join different-resolution
RSIs according to fusion levels and classes (e.g., spatial-,
spatial-spectral, and spatiotemporal-based fusion). Multisource data fusion techniques are shifting toward multialgorithm, multifeature, and multispace-time fusion. Moreover,
deep-learning methods can conduct self-sample training
with fewer labeled samples, extract key features, establish
the feature sample space, verify input data, and self-update
the feature sample space. These approaches can also extract
multifeatures from any RSIs via supervised or unsupervised
learning methods. Therefore, deep-learning techniques
have great potential for automatic UIS detection (see the
"Evaluation" section).
Additionally, applications of UIS detection can promote
the development of remote sensing-based algorithms and
models. A new relationship is proposed in Figure 13, which
clearly depicts a correspondence between UIS detection
methods and applications along with spatial resolution
and spatial scale. It outlines the major implications when
selecting reasonable methods and data to satisfy application detection goals. In Figure 13, ellipses, rectangles, hexagons, and circles respectively denote suitable spatial resolution RSIs, classification and extraction algorithms, existing
applications, and potential development. Development
trends in data sources and methods and corresponding
SEPTEMBER 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

application ranges on a spatial scale are also displayed. The
development of methods and applications has inspired researchers to propose novel techniques in UIS detection and
explore new application fields. This relationship provides a
reference for researchers to select optimal methods and appropriate RSI resolutions for specific UIS applications.
EVALUATION
The challenges and solutions of UIS detection vary with the
application requirements. Challenges will continue to emerge
in new applications (e.g., higher accuracy requirements, new
data sources, and new application scenarios), and corresponding methods will be proposed. To address new application requirements, it is necessary to consider ways to manage
challenges, choose suitable methods, realize methodological
improvements, and develop more advanced methods. With
such progress, analytical results should improve.
MANAGING CHALLENGES
Spatial resolution is a key consideration in remote sensing
data selection. Spatial resolution selection mainly depends
on the accuracy requirements, scale of the application field,
and cost of acquiring images. Generally, high-spatial-resolution RSIs are suitable for UIS detection on a local scale,
whereas medium to low spatial-resolution RSIs are suitable
for global or regional UIS detection. Data source selection
also considers weather and climate conditions in the study
areas, especially for optical remote sensing data selection
[221]. Optical RSIs are greatly affected by changes in the
external environment. In particular, the alternation of the
seasons, regional differences, and meteorological conditions exert great impacts. In areas with different climatic
zones, UIS is reflected differently in different seasons.
Moreover, the interference from bare soil should be taken
seriously. The application of auxiliary remote sensing data is
helpful for improving identification accuracy between bare
soil and UIS, such as the use of TIR images [222], night-light
data [223], and trajectory data on human activity [224]. Additionally, active remote sensing data, such as lidar and SAR
83

IEEE Geoscience and Remote Sensing Magazine - September 2019

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