IEEE Geoscience and Remote Sensing Magazine - September 2019 - 81

model uses different algorithms to extract the features of
target recognition for each data source. Then, the feature
information is analyzed and fused synthetically. The data
sources correspond to each other one by one in space and
are assigned to each other. Algorithms based on featurelevel data fusion include statistical methods and neural
networks [187], [188].
Decision-level-based fusion is a form of high-level fusion
based on image understanding and image recognition. It is
directly application oriented and provides decision support,
including feature extraction and feature recognition [189].
Major methods include the Markov random field model
technique, Bayesian rule classification theory and method,
fuzzy set theory, and expert system approach [190]. Additionally, multisource data fusion models are classified into
three types according to the image resolution and data source
characteristics of the fusion: spatial-based fusion, spatial-
spectral-based fusion, and spatiotemporal-based fusion.
SPATIAL-BASED FUSION MODEL
The spatial-based fusion approach offers advantages of
better spatial resolution to enhance the spatial resolution of object images, capturing different spatial scales
to overcome the deficiency of high spectral resolution
with low spatial resolution (e.g., SPOT 5) or high spatial resolution with low spectral resolution (e.g., IKONOS) [191], [192]. For instance, lidar point cloud and
multispectral aerial photographs were fused in UIS using an expert-systems-based method and the ISODATA
algorithm [193]. The classification accuracy and Kappa
coefficient increased by approximately 3% and 5.9%, respectively, compared with the ISODATA method alone.
Airborne SAR data (ENVISAT ASAR, TerraSAR-X) and
optical images (SPOT 5, Landsat ETM+) can be fused to
extract UIS by RF. This method can reduce confusion between bare soil and impervious surfaces, water surfaces,
and shadow areas.
SPATIAL-SPECTRAL-BASED FUSION MODEL
The spatial-spectral-based fusion model can extract spatial and spectral features simultaneously for optical RSIs.
The aim of the model is to reduce the confusion from
ground objects with similar reflectivity (e.g., LIS and
bare soil). Sentinel-1A SAR images have been combined
with Gaofen-1 multispectral images to extract impervious surfaces in Wuhan, China, using RF and Dempster-Shafer combination rules for decision-level fusion,
achieving an overall accuracy of 95.33% [194]. Research
has shown that lidar and SAR data provide additional information on landscapes to compensate for the deficiencies of optical remotely sensed data induced by environmental and weather effects. Lidar and optical data fusion
can improve extraction accuracy by approximately 3%;
SAR data and optical data can enhance such accuracy by
approximately 3.5% compared with a single optical data
source [195], [196].
SEPTEMBER 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

SPATIOTEMPORAL-BASED FUSION MODEL
The spatiotemporal-based fusion model combines images
with high spatial resolution but low temporal resolution
(e.g., IRS, 6 m and 24 days) and other images of high temporal resolution but low spatial resolution (e.g., SPOT, 250-
1,000 m and daily) to minimize adverse effects. The main
approaches with this method include those that are unmixing based [197], filter based [198], and learning based [199].
Deep-learning techniques have been proposed with this
technique to build multiple learning networks for feature
extraction. Deep CNN technology is a typical deep-learning
algorithm; it builds multilayer CNN spatiotemporal fusion
models to combine the characteristics of high temporal resolution in MODIS images and the spatial traits of Landsat
ETM+ images [200]. Fusion improves feature extraction by
automatic feature learning via multiple hidden layers.
Additionally, other data sources combined with optical
RSIs can improve the accuracy of UIS detection, such as
nighttime light data [201], [202], meteorological station
data [203], and time-series-based urban development
data [204], [205]. However, selecting appropriate data
sources and fusion algorithms for multisource data fusion
is important. Lidar and SAR data have good spatial resolution and clear geometric characteristics and are suitable
for regional or local UIS detection when fused with optical
RSIs. In particular, airborne laser scanning point clouds
data are widely used in urban feature modeling and classification [206], [207], [230].
Coherent speckle noise and side-view imaging of radar
images can also affect the accuracy of UIS detection. Moreover, UIS does not change much within a short time window, whereas other ground objects (e.g., vegetation) change
substantially with the seasons. Temporal features of images
are helpful in UIS detection in multisource data fusion,
which represents a development trend in UIS detection.
THE METHODS' ADVANTAGES AND DISADVANTAGES
To further explain the main algorithms used for UIS extraction and classification, we have identified the advantages and
disadvantages of each in Table 4. The algorithms and classifications are not meant to suggest that these algorithms are applicable only to classifications and RSIs at this scale. The listed algorithms are divided according to the RSIs of different
resolutions in the literature. Table 4 lists the major methods
and models as well as application references for UIS detection corresponding to adaptive spatial resolution RSIs. This
table also describes the advantages and disadvantages of various algorithms with different spatial resolutions. We believe
that machine-learning algorithms have great promise for UIS
detection, and an evaluation of popular machine-learning
algorithms appears in the "Evaluation" section.
A RELATIONSHIP FRAMEWORK
RSIs provide potential resource support for UIS detection
with various launched satellites and expanding applications.
The three classes of RSIs in terms of spatial resolution are
81

IEEE Geoscience and Remote Sensing Magazine - September 2019

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