IEEE Geoscience and Remote Sensing Magazine - September 2021 - 129

possible to simulate complex relationships among different
LST images so that a trained model can be used to generate
SCHR-LST [142]. Among the recently developed methods,
reconstruction and spatiotemporal fusion based on the DL
framework have already been reported.
For example, Zhang et al. [143] and Malek et al. [144]
proposed effective CNN models to recover missing data in
remote sensing images. Song et al., Tan et al., and Liu et al.
proposed novel spatiotemporal fusion models using deep
convolutional neural network [145], deep convolutional
spatiotemporal fusion network [146] and StfNet [147] to
fuse Landsat and MODIS reflectance data from different
perspectives.
Despite its proven efficiency for reflectance images, DL
has rarely been used for LST retrieval. As pioneering works,
a multiscale feature connection reconstruction CNN for
geostationary satellite LST images with large missing regions
[148] and STTFN [116] have been proposed. These
methods provide a new framework and advanced capabilities
for obtaining spatially continuous and high-resolution
LSTs or other remotely sensed data products. In the future,
we will apply more DL studies to LST research owing to the
availability of sample data of high quality.
However, imperfect knowledge of remotely sensed
LSTs and low computing efficiency are key factors limiting
the wide application of DL to reconstruction and
spatiotemporal fusion. However, the available LST data
will continue to grow and improve with the development of
large data methods and further advances in remote sensing.
In addition, more advanced processing strategies (i.e., parallel
computing and graphics processing unit acceleration)
and advanced cloud-based geospatial processing platforms
(e.g., Google Earth Engine) have been put forward
[149], [150], thereby addressing problems with computing
efficiency. All of these developments will significantly
improve the application potential of remotely sensed LST
products, which is the ultimate goal of LST reconstruction
and spatiotemporal fusion.
COMBINING DATA ASSIMILATION WITH
RECONSTRUCTION AND SPATIOTEMPORAL FUSION
Data assimilation provides continuous information on
variables and locations without direct observation by using
proxy data, thereby filling gaps between sparsely distributed
observations [151]. However, input uncertainty easily
leads to the accumulation of errors by the model performing
the assimilation, and substantial surface heterogeneity
limits its application, e.g., over urban areas.
Remote sensing data can provide accurate spatial information
and do not accumulate errors over time. Using
data assimilation, LSTs can be combined organically with
land surface process models to optimize the simulated
TABLE 7. A SUMMARY OF APPLICATION CONDITIONS, FEATURES, AND DATA SOURCES OF RECONSTRUCTION
AND SPATIOTEMPORAL FUSION METHODS.
METHOD
APPLICABLE CONDITIONS AND FEATURES
Spatial reconstruction Homogeneous landscapes with a small number of missing pixels
High efficiency
Hypothetical LST values provided under cloudless conditions
Temporal
reconstruction
Spatiotemporal
reconstruction
PWM-based
reconstruction
SEB-based
reconstruction
Weighted functionbased
fusion
Unmixing-based
fusion
Hybrid methods
fusion
Learning-based
fusion
Regular temporal variation with larger missing regions
Requires multitemporal data
Provides hypothetical LST values under cloudless conditions
Applicable to most cases of missing LST data
High accuracy
Provides hypothetical LST values under cloudless conditions
Requires corresponding microwave data
Estimates the LST values under actual cloudy conditions
Requires consideration of differences in resolution, subsurface temperature,
and LST
Makes some assumptions on environmental variables and requires specific
meteorological observations
Estimates the LST values under actual cloudy conditions
Has gained considerable popularity and applications
Requires cotemporal (reference time) pairs of fine- and coarse-spatialresolution
LST data in addition to coarse-spatial-resolution LST data
at the prediction time
Requires coarse-spatial-resolution LST data at the prediction and reference
times as well as high-spatial-resolution images near the predicted date
Integrates two or more techniques and contains weighted function- and
unmixing-based methods
High accuracy
A promising approach with powerful nonlinear representation ability
between input and output images
Relies heavily on training samples and several customized model parameters
APPLICABLE LST DATA SOURCES
LST TIR-based from polar-orbiting sensors,
such as MODIS and Landsat [60]
Polar-orbiting and geostationary satellites,
such as GOES, SEVIRI, FY-2F, and others
[62], [72]
Both polar-orbiting and geostationary
satellites [74], [75]
AMSR-E/2 data used to reconstruct MODIS
LST [24], [79], [85]
PWM data also have the potential to
reconstruct geostationary satellite LST
MODIS LST is popular [81], [88], [90]
Applicable to almost all LSTs from TIR
sensors
Fusion of MODIS and Landsat is popular
[100], [103], [104]
SEVIRI or GOES LST also used as
coarse-spatial-resolution data [17]
Fusion of MODIS and ASTER was
tested [111]
Fusion of MODIS and Landsat was tested [50]
FY-2F LST also used as coarse-spatial-resolution
data [105]
Fusion of MODIS and Landsat attempted
[115], [116]
Fusion data from satellite and ground
sensors show potential
SEPTEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
129
IEEE Geoscience and Remote Sensing Magazine - September 2021

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