IEEE Geoscience and Remote Sensing Magazine - December 2021 - 81

According to the processing unit, deep learning-based detectors
are divided into patch- [116], [130], pixel- [110], [117],
and object-based [127], [140] varieties. For a patch-based
change detection task, a sliding window with a fixed size is
used to divide the study area into a series of patches, and each
patch is assigned a label by the detector. In this way, each pixel
in the patch is assigned the same label. Consequently, rough
location-not fine-grained-boundary-of-change information
is obtained. However, patch-based change detection can
reduce the influence of spatial misalignment to some extent
in VHR change detection. Since patch-based deep learning
networks view each patch as the change detection unit and
encode each patch as a set of feature maps with coarser spatial
resolutions, the spatial misalignment of these feature maps
becomes smaller, and some errors of spatial alignment are
therefore avoided in a change detection task. In other words,
when regarding a patch as the change analysis unit, only a
very large misalignment can cause an unchanged image
patch to be identified as a changed one, and a small misalignment
can be tolerated. Several important issues should be noticed
for the patch-based method, such as the oversmoothing
of results and the selection of the patch size.
The multiscale strategy [135] may be appropriate for addressing
these issues, but it inevitably leads to larger computation
burdens. Pixel-based methods usually employ
semantic segmentation architectures to predict pixel-wise
change detection results [33]. Specifically, in semantic segmentation
architectures, after extracting abstract semantic
information through multilayer encoding (e.g., convolution
layers), a series of operations, e.g., interpolation, deconvolution,
and upsampling, is used to progressively decode
semantic information into feature maps that have the
same spatial resolution as the input images. Unlike traditional
pixel-based change detectors that suffer from misregistration,
viewing angle differences, and occlusions, deep
learning methods can predict pixel-wise change detection
with a highly semantic abstraction of the spatial context.
However, object boundaries are often blurred in the change
detection results, as up-sampling layers reconstruct the appearance
but not the shape of objects. To cope with this
issue, better networks are designed. UNet++, for example,
combines nested features to preserve change region boundaries,
considering that shallow layers are better able to capture
spatial details [110].
Object-based deep learning methods are also considered
for change detection [127], [140]. A simple approach is
to adopt object-based segmentation in the pre/postprocessing
step, as shown in [140]. On the other hand, object information
can be also considered during the training process
by adding object-wise loss terms [127]. However, issues related
to conventional multitemporal image segmentation,
such as oversegmentation, undersegmentation, and " sliver
objects " caused by misregistration, remain unsolved. In the
future, object-based detectors need to generate semantic
segments and establish spatial correspondence between
multitemporal segments.
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
The types of characteristics most often used for each criterion
(i.e., the learning strategy, fusion strategy, network
model, and processing unit) in VHR change detection are
summarized in the following:
1) For the learning strategy, supervised learning is the most
widely used method for VHR change detection. However,
the great amount of labor required to collect a large
number of training samples becomes a bottleneck, especially
for deep network models, which leads to increasing
attention for other learning strategies.
2) Late and early fusion strategies have their own strengths
and weaknesses in representing multitemporal features
and their differences, and hence hybrid fusion is sometimes
chosen.
3) Among various network models, CNNs are the most commonly
considered, and they are coupled with other networks,
i.e., hybrid models, for instance, CNN-RNNs [111].
4) As for the processing unit, most studies consider patchand
pixel-level models. Patch-level detectors are more
tolerant to spatial misalignment, but pixel-based ones are
more appropriate for identifying fine-grained changes.
APPLICATIONS OF VHR CHANGE DETECTION
VHR image change detection is widely used in a large number
of practical scenarios. A series of representative applications
is the focus of this review, including the monitoring
and change detection of 1) land cover and land use,
2) buildings, 3) vegetation, 4) crops, 5) lakes and wetlands,
6) ecosystem services, and 7) impervious surfaces.
LAND COVER AND LAND USE CHANGE DETECTION
Compared to coarse- and medium-resolution images, VHR
images can reveal detailed and subtle intraurban change
information [141]. Specifically, urban change detection by
combining multiple features (e.g., object-based spectral,
shape, and texture attributes) was presented in [142], where
changes to detailed urban objects, e.g., buildings, roads, and
playgrounds, can be detected. Huang et al. [21] identified
pixel-level change transitions in 2012-2013 using Ziyuan-3
orthographic images, and the experimental result is presented
in Figure 11. It can be seen that, even in the one-year period,
small-scale changes extensively occurred in the urban
area of Wuhan, China. For instance, fine-scale urban land
cover transitions caused by pond infilling, building demolitions,
building construction, weed growth, and site preparation
can be observed. In [143], changes in detailed land
cover classes, including bright roofs, gray roofs, tile roofs,
brown fields, dark asphalt, light asphalt, and so on, were
analyzed using IKONOS and GeoEye-1 images.
As for land use change detection, Wu et al. [108] interpreted
change transitions, e.g., from sparse housing to industrial
areas, by combining spectral and SIFT features. In [144], land
use maps of Shenzhen (a highly dynamic and developed
megacity in China) were generated in 2005 and 2017 based
on VHR satellite data. As demonstrated in Figure 12, detailed
land use categories, including residential, commercial,
81
IEEE Geoscience and Remote Sensing Magazine - December 2021

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