IEEE Geoscience and Remote Sensing Magazine - December 2021 - 72

angular second moment and entropy), and statistical (e.g.,
the mean, variance, and correlation) attributes within local
windows [39], [40]. The LBP, an ordered set of binary
comparisons of pixel values between the central pixel and
its neighboring ones, is invariant to monotonic grayscale
change [41]. The PSI aims to measure the length of direction
lines, which are extended based on gray-level similarity
along a series of directions [42]. Some representative examples
for VHR change detection using statistical textures
are briefly introduced in the following.
Tan et al. [43] adopted the GLCM in an automatic change
detection method to consider the variation information of
direction, distance, and amplitude in images. Li et al. [44] applied
the local similarity of GLCM textures to detect changes
and demonstrated that this kind of feature was robust against
both noise and spectral similarity. Peng and Zhang [45] used
the LBP for change detection from Gaofen-1 imagery, and
both qualitative and quantitative analyses demonstrated
the effectiveness of the proposed approach. Zhang et al. [46]
identified building change types, i.e., new construction,
demolition, and reconstruction, by using LBP features
and obtained satisfactory change detection results with a
high detection accuracy and precise structure boundaries.
Liu et al. [47] proposed a line-constrained shape feature, a
modified version of the PSI, for building change detection,
and the results showed the approach's advantage in individual
building change detection in a lightly populated region.
Structural textures, e.g., morphological profiles (MPs)
and attribute profiles (APs), facilitate the investigation
of the geometries, shapes, and edges of regions, with the
convex and concave components being erased so that the
geometric information of relevant structures is preserved
and unimportant details are attenuated [48], [49]. MPs
and APs have proved to be effective in VHR change detection
since they can simplify results and reduce noise components
(e.g., spectral variations) [48], [49]. For instance,
Liu et al. [50] took the geometrical structure of change targets
into account using MPs. In addition, the morphological
building index (MBI) [36], which is defined as differential
MPs with linear structural elements, has been extensively
used in VHR change detection in urban areas since it can
highlight bright and high-contrast structures, mostly consisting
of buildings, in remote sensing images. For example,
Huang et al. [51] proposed an automatic building change
detection framework based on the MBI. Experimental results
showed that the proposed method outperformed supervised
classification via a support vector machine (SVM).
In addition, point and line features, for instance, Harris
[52] and scale-invariant feature transforms (SIFTs) [53], can
improve the discriminability of man-made objects, such as
buildings, roads, and cars, by describing corners and edges,
therefore improving results.
Model-based textures, e.g., Markov random fields (MRFs)
and fractal models, aim to represent textures through stochastic
processes [54]. MRF models present spatial context through
a graph-based image representation, where the nodes and
edges of the graph express pixels and their relationship with
connected nodes, respectively. Fractal models can depict texture
roughness and complexity by capturing self-similar and
self-affine patterns [55]. A number of MRF-based methods
have been proposed to deal with VHR image change detection
[56]-[60] because of their ability to describe local spatial
TABLE 2. A SUMMARY OF THE FEATURES USED FOR VHR IMAGE CHANGE DETECTION.
CATEGORY SUBCATEGORY
DESCRIPTION
Textural
features
Statistical
Structural
Model based
Deep
features
Describe the relationships among the gray levels of
local windows
Investigate the geometry, shapes, and edges of
regions
Obtain coefficients from the model describing the
relationships among the local image neighborhood
Transform based Capture local structures in a transformed space
Autoencoders
Convolutional
neural networks
Objectbased
features
First
level
Second level
Angular
features
Third level
Implicit
Explicit
Extract mid- and high-level abstract features by
interleaving convolutional and pooling layers
Radiometry, geometry, and texture for each image
object
Relationships between two image objects, e.g.,
adjacency and proximity, and relationships with
neighboring objects
Spatial arrangements of multiple objects
Orthographic images and DSMs
Quantify the differences contained in multiangle
images, such as angular difference features
Learn efficient encoding through the optimization
of a series of criteria
Complex training and
parameter tuning,
" black-box " nature, high
computational burden,
overfitting, and so on
Determination of appropriate
segmentation
parameters and uncertainties
of the segmentation
results
Availability of multiangle
images
Gaofen-2
[66], [77]
and Google
Earth images
[66], [76]
QuickBird
[88], [89]
CHARACTERISTICS
Edge effect, difficulty of
identifying parameters
SENSOR
QuickBird
[48]-[53]
REFERENCES
[43]-[47]
[48]-[53]
[56]-[61]
[63], [64]
[70]-[73]
[66], [67],
[75]-[78]
[85], [88],
[89]
[91], [92]
[95]
Ziyuan-3 [2],
[21]
[21], [98],
[99]
[2]
72
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2021
IEEE Geoscience and Remote Sensing Magazine - December 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2021
