IEEE Geoscience and Remote Sensing Magazine - December 2019 - 10

Cresson and Saint-Geours [26] proposed a global harmonization method by solving a quadratic programming optimization problem. This method can implement multiple
remote sensing images simultaneously without any given
reference image. However, the methods introduced in the
following are usually based on a reference image. The main
methods of radiometric normalization can be classified
into global models, local models, and combined models.
GLOBAL MODELS
Global-model-based methods assume that the radiometric mapping relationship between the source and target
images can be represented by a global linear or nonlinear transform, as shown in (1) [27]. The source image is
the image selected as a reference, and the target image is
the image whose intensity is to be corrected. The overall radiometric consistency is ensured by the following
global model:
I *1 = f (I 1),

(1)

where I 1 is the target image, I *1 is the corrected target image, and f ($) is a linear or nonlinear function denoting the
mapping relationships of all of the bands. Global-modelbased methods can be further grouped into pixel-to-pixel
and region-to-region methods.
PIXEL-TO-PIXEL METHODS
Pixel-to-pixel methods directly model the radiometric
mapping relationship using the intensity values of the corresponding pixel pairs. When the objects in the overlapping
areas are not changed or the images are captured very closely in time, the relationship of the pixel pairs can be regarded as linear. In this situation, linear regression [25], [28]
and least-mean-square (LMS)-based transformation [29],
[30] for all of the pixel pairs are two effective approaches.
In addition, Chen et al. [31] proposed a method based on iteratively reweighted, multivariate alteration detection transformation and orthogonal regression [32] to reduce the error of radiometric normalization: they achieved radiometric
consistency by considering the effect of the normalization
path on the normalization coefficients.
Usually, all pixels in the overlapping areas are used to
build the relationship. However, not all pixels always satisfy the linear assumption. Thus, the characteristic pixels
that meet the linear assumption in the overlapping areas
need to be carefully selected. For example, Yong et al. [33]
applied band-specific principal component analysis to select the characteristic pixels. Zhang and Georganas [34]
selected the principal regions using an intensity histogram
to construct the transform matrix based on the average intensity values. This method is very fast, but it is not known
how the accuracy of the registration affects the radiometric normalization. Radiometric inconsistency can also be
corrected according to the imaging mechanism. Litvinov
and Schechner [35] corrected radiometric mismatch by
10

estimating the radiometric response and camera nonuniformity simultaneously, based on a computer-vision tool
and the physical process of the imaging system. The results
they obtained were very satisfactory, indicating that mosaicking can be successfully achieved without resorting to
any type of feathering method.
Pixel-to-pixel methods utilize the mapping relationship
directly derived from the pixel pairs to correct the radiometric differences. These are the basic methods of radiometric normalization, and they usually obtain satisfactory
results for consistent radiometric differences. However, in
most cases, they are very sensitive to the accuracy of the
image registration. These methods can thus achieve a good
effect when registration accuracy is high and radiometric
difference is consistent.
REGION-TO-REGION METHODS
The region-to-region methods utilize the statistical information (e.g., mean, standard deviation, and variance) of
the intensity in the overlapping areas to construct the radiometric mapping function [36]. Compared with pixel-topixel methods, region-to-region methods indirectly model
the mapping relationship from pixel pairs. An advantage
of this method is that the pixel pairs are not required to be
strictly aligned. When the objects in the overlapping areas
are changed, especially for multitemporal remote sensing
images, region-to-region methods are more effective than
pixel-to-pixel methods. The representative methods include diagonal-matrix transformation models [37], [38],
histogram matching [39]-[42], moment matching [43],
[44], Wallis transformation [45], and quadratic programming color balancing [26].
Among region-to-region methods, the diagonal-matrix transformation models are very basic and simple,
and they use the mean intensity of the neighboring images to calculate the relationship. For example, Tian et al.
[37] proposed a six-parameter diagonal model to compensate for the radiometric differences by spectral transformation between images. In their view, the basis of this
approach is that the reflected light depends on the spectral properties and the illumination angle on the surface.
Generally speaking, the diagonal models are suitable for
remote sensing images with ordinary scenarios or a low
spatial resolution.
The histogram-matching-based methods make the histogram in the overlapping areas of the target image similar
to that of the source image. These methods assume that
the radiometric mapping function has no particular parameters, and most of them apply a lookup table to directly record the mapping relationship of the source and
target images [27]. Generally, the lookup table is constructed from the joint histogram of the image features
or pixel pairs in the overlapping areas. Interestingly, Xie et
al. [46] proposed global optimization to realize intensity
consistency, guided by an initial solution of the histogram
extreme-point matching strategy. Based on the histogram,
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2019

IEEE Geoscience and Remote Sensing Magazine - December 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2019