IEEE Geoscience and Remote Sensing Magazine - June 2023 - 46

transformation features; and electromagnetic features,
such as the emitter center frequency, pulse width, and
pulse repetition interval.
Target-association algorithms between imaging satellites
can be realized using image-matching methods. The
key step of this method is feature matching that uses feature
similarity measurements of the target images to build the
matching relationship. The image features used for feature
matching include both single- and group-target features.
Single-target features are primarily used for target associations
in high-resolution satellite remote sensing images. Lei
et al. [143] proposed a target association algorithm based
on multiscale autoconvolution features and association
cost matrix for optical satellites. Group-target features are
primarily used for the target association of medium- or
low-resolution satellite remote sensing images. Group targets
typically appear in a relatively fixed formation, and the
membership and number of targets are typically related to
specific operational tasks. Therefore, the group-target position
can be regarded as a point set on the plane and each
target in that group can be regarded as a point in the point
set. Next,
the multitarget association problem is transformed
into a matching problem between the two point
sets. Tang and Xu [144] proposed a target association algorithm
based on the Kalman filter and the Dempster-Shafer
(D-S) theory for multiple remote sensing images.
Target association based on image matching focuses on
the target feature design and feature similarity measurement.
Another type of association method is based on
point features, mainly designed for GEO staring imaging
satellites and LEO video satellites, which use motion status
parameters and image features as association factors. The
type of methods uses filtered and predicted target motion
status parameters from image frames for the first-step association,
and they use image feature similarity to correct
the association result as the second-step association. Lei
[145] proposed a remote sensing image multitarget tracking
algorithm based on ROI feature matching and a remote
sensing image multitarget association algorithm based on
multifeature fusion matching, which overcome the bias of
kinematic feature matching error through image features
and the ambiguity of image matching recognition through
motion state parameters. Yang et al. [146] proposed a satellite
video target correlation tracking algorithm based on a
motion heat map and local saliency map. Wu et al. [147]
proposed a satellite video target correlation tracking algorithm
that combines motion smoothing constraints and
the grey similarity likelihood ratio. Liu et al. [148] proposed
a multilevel target association method using different features
at different levels for the collaborative observation of
LEO and GEO satellites.
Target association algorithms between nonimaging and
video-imaging satellites, such as AIS, SIGINT, or ELINT,
mainly combine position and attribute information based
on fuzzy mathematics [149] and D-S [150]. Some algorithms
use group-formation topological characteristics as
46
association features that can be considered point-pattern
match problems. For target association between imaging
satellites and nonimaging satellites, Zeng [151] proposed
several target association algorithms based on formation
target structure features, hierarchical matching, formation
target attribute features, the attribute and structure of formation
targets, and multisource features in multisource sequence
data. Lu and colleagues [152], [153] proposed target
association algorithms based on point-pair topological features
and spectral matching, point-pair topological features
and probabilistic relaxation marking, and D-S evidence
combinations based on topology and attribute features.
CURRENT STATUS OF MULTISATELLITE DATA FUSION
Information fusion algorithms can be classified into three
levels based on the information abstraction level: data, features,
and decisions. Feature-level fusion can maximally
retain most of the information of the original data but can
also greatly reduce the redundancy of multisource data.
Current research mainly focuses on data- and decisionlevel
fusion, whereas research on feature-level fusion is
relatively scarce. However, feature-level fusion reduces
the dimensions of the original feature space, eliminates the
redundancy between feature representation vectors in the
original feature space, and maintains the entropy, energy,
and correlation of invariant feature data after dimension
compression. The fused features can substantially describe
the nature of the target, which is conducive to further target
recognition. The following analysis mainly focuses on
feature-level fusion algorithms.
Conventional feature-level fusion algorithms include
serial and parallel fusions [154]. Serial fusion methods directly
concatenate multiple feature vectors into one feature
vector and then apply a dimensional reduction to obtain
the fused feature vector. Parallel fusion methods combine
two feature vectors into a single feature vector using complex
variable functions. Although the above two methods
can retain raw data information to some extent, the dimension
and redundancy of the fused feature remain high because
the complementarities and redundancies between
the original multiple features are not fully utilized.
Feature extraction and transformation are also considered
as feature fusion methods. Feature extraction obtains
a fused feature by selecting the most effective feature from
multiple original features using serial or parallel strategies.
Feature transformation obtains a new fused feature
through a linear or nonlinear mapping function of the
original features and can still be considered as a type of feature
extraction method. Some feature-level fusion methods
based on multivariate statistical analysis theory have been
proposed to solve the problem of serial fusion and parallel
fusion methods not being able to use the interrelationship
between multidimensional features. Feature fusion methods
based on canonical correlation analysis (CCA) obtain
the fused feature by building a correlation criterion function
between two feature vectors, calculating the projection
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2023
IEEE Geoscience and Remote Sensing Magazine - June 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2023
