IEEE Geoscience and Remote Sensing Magazine - March 2022 - 55

have various accuracies for a specific data set in terms of
varying indicators. This visual performance conclusion
well supported the quantitative comparisons in Tables 3-5.
DISCUSSION
To obtain an overview of the development of LCCD with
VHR optical images, we statistically analyzed the articles
referred to. In accordance with the perspective and focus of
these articles, we discussed LCCD with VHR optical images
from the following viewpoints: technology, application,
data sets, and accuracy assessment. The details of this quantitative
analysis are presented in the following sections.
TECHNIQUES
As mentioned in the previous section, here, we dichotomized
the major LCCD methods into four groups in terms
of different manners of utilizing spatial contextual information:
regular neighboring context [45]-[48], multiscale
neighboring context [5], [51]-[53], adaptive neighboring
context [27], [60], [61], and other methods borrowed from
computer vision. As shown in Figure 9, the regular and
multiscale context-based approaches account for 31.7%
and 45.2% of the total articles, respectively. However, the
adaptive context-based approaches account for only 14.6%
of the total quantity. The discrepancy could be due to the
fact that more than two parameters are required by an algorithm
that explores adaptive context [27], [63], [148].
Moreover, the optimal parameter's setting of an adaptive
context exploration algorithm for a specific data set remains
a challenge [6], [27], [60]. In practical applications, trial-anderror
approaches are usually adopted for identifying the optimal
set of parameters for a given data set and a specific algorithm.
Hence, multiscale context-based LCCD approaches
play a dominant role in the technical development of LCCD
with VHR images. Moreover, adaptive context requires more
attention in terms of obtaining the optimized parameters
and improves the degree of the model automatically for the
practical application of LCCD with VHR optimal images.
APPLICATIONS
LCCD with VHR images has been widely used in practice
[33], [41], [149]. In this section, we refined the 356 articles
from the viewpoint of application for thematic analysis.
Figure 10 illustrates that many LCCD applications with
VHR optical images focus on urban LCCD.
Based on these observations, 92 study areas were located
in the urban region, accounting for 30% of the total
quantity. In addition, building change detection played
an important role in change detection with VHR optical
images, accounting for 23.8% of the 307 applications in
the 356 articles. However, the application of such methods
to disaster change detection, such as earthquake and
landslide inventory mapping, has not gained sufficient
attention, probably because of the difficulty of acquiring
pre- and postevent images for natural disaster events.
Furthermore, coregistration between bitemporal
images
MARCH 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
becomes challenging because land cover disasters, such as
landslides and earthquakes, usually result in the deformation
of ground elevation.
DATA SETS
Data sets are important in investigating the performance
of LCCD algorithms. If an algorithm adapts to more kinds
of data sets with different resolutions, then it may have
better adaptability and robustness [73]. Thus, we reviewed
the data sets mentioned in the selected articles in terms of
data sourcing and size of the study areas. Figure 11 demonstrates
that approximately nine kinds of data, including
QuickBird, aerial orthophotos, WorldView, IKONOS,
GaoFen, SPOT, GeoEye, UAV images, and others (ZY-3, BeiJing-2,
and Pleiades), were mentioned in the selected articles.
QuickBird was the most widely used data set, accounting
Regular
Context
113 (31.7%)
Multiscale
Context
161 (45.2%)
Others
30 (8.4%)
Adaptive
Context
52 (14.6%)
FIGURE 9. The statistics for LCCD as the subject of articles on VHR
optimal images published in 2001-2019 from the viewpoint of
techniques.
Water Body CD
10 (3.3%)
Vegetation CD
20 (6.5%)
Building CD
73 (23.8%)
Earthquake CD
9 (2.9%)
Urban CD
92 (30%)
Landslide
Inventory
Mapping
22 (7.2%)
Others
62 (20.2%)
FIGURE 10. The statistics for LCCD as the subject of articles on
VHR optimal images published in 2001-2019 from the viewpoint of
applications. In the 356 articles, 307 experiments mentioned the
characteristic of studying the area.
55
Map Updating CD
19 (6.2%)
IEEE Geoscience and Remote Sensing Magazine - March 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2022
