IEEE Geoscience and Remote Sensing Magazine - December 2021 - 77

set, which is available at http://irsip.whu.edu.cn/resources/
resources_en_v2.php. It includes both urban and rural scenes
with diverse and complex change types, and, moreover,
it considers seasonal and illumination influences. These
characteristics enable the MVCD to function as a challenging
change detection data set. A comparative analysis
between different attributes, including the GLCM [39], AP
[49], CT [62], MABI, [101], object-wise GLCM (GLCM-Obj)
[87], and deep features [67], has been carried out. Specifically,
the change intensity map was obtained by CVA, and
the threshold for each feature was determined based on receiver
operating characteristic curves to achieve a balance
between commission and omission errors [65]. Qualitative
and quantitative experimental results are provided in
Figure 8 and Table 3, respectively. The spectral feature fails
to detect changes between spectrally similar classes (e.g.,
bare soil and buildings), and unchanged objects with spectral
variation are incorrectly detected as changed ones. The
GLCM, AP, and CT can depict textural changes, e.g., the
spatial distribution of the gray value, geometry, and local
details. Among them, the CT gives more complete changed
regions, and the AP produces more false alarms. The MABI
emphasizes building changes, but it is not sensitive to other
variations (e.g., soil, vegetation, and roads), which therefore
leads to a large omission error. The GLCM-Obj generates
smoother results with smaller omissions but larger
commission errors than its pixel-wise version. Deep CVA
outperforms the other methods, but false alarms caused by
shadows and seasonal effects can be still observed.
CHANGE DETECTORS
VHR change detectors can be categorized as algebra-, transform-,
and machine learning-based indicators. CVA is one of
the most widely used algebraic approaches, and it is carried
out by measuring the difference among bi-temporal multifeature
vectors to derive a change vector for VHR images
[67], [104], [105]. Transform-based methods, such as principal
component analysis [106] and multivariate alteration
detection [107], attempt to suppress no-change areas and
emphasize change information in the transformed feature
space. In the machine learning community, change detection
is often viewed as a classification problem. In conventional
classification-based VHR change detection, spectral-
spatial feature extraction and detectors (e.g., SVMs [108] and
the random forest [65]) are separately implemented. The
recent hot spot, i.e., deep learning, can integrate these two
operations in a joint learning framework, which is therefore
very promising for VHR change detection [109], [110].
Deep learning-based change detectors can be grouped
in terms of different criteria, including learning and fusion
strategies, network models, and processing units (Table 4).
We first discuss learning strategies. On the basis of a
large amount of annotated data, supervised deep learning
methods can capture semantic changes, and hence they
1
2012
2013
2014
MBI
Harris
Pantex
2015
2016
4
2
3
2018
(a)
Reference Data
Fused
Planar Features
Fused
Vertical Features
(b)
Planar
Vertical Features
FIGURE 7. Experimental results for the automatic monitoring of newly constructed building areas (NCBAs) using planar (i.e., MBI, Harris,
and Pantex) and vertical (MSI, nDSM, and MABI) features [2]. (a) Multitemporal Ziyuan-3 images. (b) NCBAs and their change timing.
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
77
2017
MSI
nDSM
MABI
2013
2014
2015
2016
2017
2018
Non-NCBAs
http://irsip.whu.edu.cn/resources/resources_en_v2.php http://irsip.whu.edu.cn/resources/resources_en_v2.php
IEEE Geoscience and Remote Sensing Magazine - December 2021

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