IEEE Geoscience and Remote Sensing Magazine - September 2021 - 128

QUANTITATIVE EVALUATION INDICES
It is significant to quantitatively evaluate the accuracy of
the reconstruction and fusion results. Numerous quantitative
indicators for evaluating reconstructed and fused
results have been adopted [5], [17], [42], [78], [134]. The
main evaluation indicators include the RMSE, MSE, MAE,
average absolute difference, mean error, CC, CD, SSIM, and
standard deviation.
It should be noted that, in this study, peer-reviewed research
papers are considered; it is beyond the scope of this
article to validate all of the findings of other researchers.
However, experimental results from significant algorithms
and their quantitative evaluation are provided here. Based
on our evaluation results and the published literature, we
conclude that, in addition to the reconstruction and fusion
methods themselves, the validation results depend significantly
on LST data as well as the distribution characteristics
and number of missing LST pixels, degree of heterogeneity
over the study area, and observation time [8], [78]. It is
not unreasonable to assess the quality of various methods
based on their individual validation results.
FUTURE PROSPECTS
Despite the achievements described, methods for obtaining
SCHR-LST are currently hot research topics. However, LST reconstruction
and spatiotemporal fusion are complicated and
inherently ill-posed inverse problems; thus, research is still
needed and provides the potential for further development.
In this section, we list and discuss several topics that appear
promising for improving SCHR-LST retrieval from spacebased
measurements.
EXPLOITATION OF GEOGRAPHICAL LAWS AND
SIGNAL PROCESSING TECHNIQUES
Geographic surface parameters (e.g., LST) occur in the realm
of space and time and, thus, follow geographical laws describing
the behavior of their spatiotemporal autocorrelation, heterogeneity,
and spatial similarity. The first two geographical
laws state that geographic environmental variables 1) are
spatiotemporally correlated with themselves and 2) vary in
space and time. The recently proposed third law of geography
concentrates more on the similarity of the geographic
configuration (SGC) of locations [135]. Spatial prediction of
the geographic environmental variables can be conducted
on the basis of SGCs between a sample and its prediction
point. However, signal processing techniques are usually
employed for establishing the global numeric relationships
among variables with or without insufficient consideration
of geographical laws.
Nearly all current reconstruction and spatiotemporal
fusion methods originate from signal processing. These
approaches perform well over small and homogeneous
geographic areas without significant land use or land cover
changes. However, when facing large and complex geographic
areas, auxiliary geographical information should
be utilized. LST is a typical geographical parameter; thus,
128
the integration of geographical laws and signal processing
techniques into SCHR-LST algorithms appears promising.
INTEGRATION OF PHYSICAL PROPERTIES
AND SIGNAL PROCESSING TECHNIQUES
Although input multitemporal and multisensor LST products
are retrieved with physical models, obtaining reconstructed
and fused LSTs often involves filtering, interpolation, regression,
variational processing, and sparse representation
methods. Previous studies have shown that the accuracy of
LST reconstruction differs significantly between nighttime
and daytime and varies with seasons [78]. Physical properties,
such as high-dynamic-change characteristics, DTC,
and ATC, might explain these results better than individual
LST values.
Furthermore, almost all reconstruction and spatiotemporal
fusion methods are implemented on LST images
from multitemporal or multisensor products. However, as
reviewed by Li et al., LSTs can vary between the viewing
zenith angle and acquisition time (local solar time) [6]. The
differences in LST measured at the nadir and off-nadir can
be up to 5-10 K for different land cover types [6]. Because
many polar-orbiting satellite sensors scan Earth's surface
in the cross-track direction, their viewing zenith angle varies
significantly (i.e., −65° to +65°); therefore, LST values
of different pixels in the same orbit cannot be compared
[136]. This effect must also be considered for multisource or
multitemporal LST products.
In addition, LST data retrieved from the same sensors
could be incomparable if their variation in observation
time is significant [137], [138]. This effect is even more obvious
in LST data from different satellites acquired at varying
times. Inevitably, angular and temporal dissimilarities
pose great challenges to reconstruction and spatiotemporal
fusion. Therefore, it is necessary to integrate physical properties
(i.e., DTC and ATC modeling as well as angular and
temporal normalization) into reconstruction and spatiotemporal
fusion methods.
NEW PROCESSING FRAMEWORK FOR
RECONSTRUCTION AND SPATIOTEMPORAL FUSION
It is well known that satellite-based LSTs change rapidly
in space and time [21]. The traditional signal processing
framework (e.g., filtering, interpolation, regression, variational
processing, and sparse representation) assumes
that the input data and output results are linear or simple
nonlinear transformations. It is difficult to conduct comprehensive
feature mining for a complex nonlinear transformation
process, nonstationary characteristics (such
as high-dynamic-change characteristics), and large-scale
differences between high- and low-resolution LST images.
In recent years, DL has gained the attention of the remote
sensing community and been used for various image analysis
and classification problems. In particular, overviews of
DL for data fusion [140], [141] and reviews on DL in environmental
remote sensing [142] have been published. It is
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE SEPTEMBER 2021
IEEE Geoscience and Remote Sensing Magazine - September 2021

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