IEEE Geoscience and Remote Sensing Magazine - September 2021 - 123

Hereafter, this approach is referred to
as the Long method. Although the SEBbased
technique includes multitemporal
reconstruction and correction with
SEB [88], Long et al. compared the LST
estimated only from the multitemporal
reconstruction with their results because
the accuracy of the corrected LST
was not improved [91].
The mean absolute error (MAE),
TABLE 4. A QUANTITATIVE EVALUATION OF THE RECONSTRUCTION METHODS
FOR CLOUDY PIXELS.
COMPARISON A
INDICATORS
MAE (ºC)
RMSE (ºC)
CD
RMSE, and coefficient of determination (CD) of the LSTs
obtained from the PMW-based and Long methods were
compared with in situ LSTs (comparison A), as shown in
Table 4. All three indicators in the Long method were better
than those in the PMW-based approach. Similarly, the
three indicators of LSTs obtained from the Long method
were better than those from the SEB-based technique
when compared with in situ LSTs obtained at other times
(comparison B).
APPLICATIONS AND LIMITATIONS OF
PMW- AND SEB-BASED METHODS
The primary applications and limitations of PMW- and
SEB-based approaches are discussed in this section. PMWbased
techniques are suitable for LST retrievals in all-sky
(i.e., cloudless and cloudy) conditions and can provide
useful information for reconstructing the cloudy LSTs of
TIR sensors. However, LSTs estimated from PMW have
limitations in accounting for the spatiotemporal variability
of microwave emissivity [92] owing to variation in land
surface types. Moreover, subsurface temperatures retrieved
from PMW differ from LSTs (also known as skin temperature)
obtained from TIR sensors and require conversion to skin
temperature [93].
Another limitation is that the resolution of the PMW data
(e.g., 25 km) is significantly lower than that of TIR data (e.g.,
100 m). These differences require consideration
when fusing TIR and PMW
LST products, such as recovering the
LST at the spatial resolution of TIR
data when a microwave pixel is partly
cloudy. Therefore, high-precision, physics-based
retrieval algorithms; effective
conversion models; and an appropriate
downscaling algorithm are urgently required
in PMW-based reconstruction
methods [24], [83]-[85].
SEB-based techniques provide a
t1
Spatiotemporal
Fusion
t0
STARFM-Like
Methods
new approach for reconstructing the
cloudy LSTs of TIR sensors without
considering the differences in spatial
resolution and conversion between
subsurface temperature and LST. However,
these methods require valid
NPs from spatial, temporal, or spatiotemporal
techniques [87], which are
Fused LST
t0
t2
PMW-BASED
METHOD IN [24]
5.96
6.89
0.8
LONG
METHOD [91]
2.31
3.01
0.91
COMPARISON B
LONG
METHOD [91]
2.49
3.24
0.93
SEB-BASED
METHOD IN [88]
3.85
5.16
0.81
difficult to obtain in humid regions with consecutive rain
events. Furthermore, these approaches make assumptions
about environmental variables and require specific meteorological
and hydrological observations to calculate LST errors
between cloudless and cloudy areas, thereby introducing
more error [80].
In summary, the PMW- and SEB-based methods can be
used to effectively estimate LSTs under all-weather conditions
compared with the spatial, temporal, and spatiotemporal
reconstruction approaches. However, their limitations hinder
their further application.
SPATIOTEMPORAL FUSION FOR HIGH RESOLUTION
By far, the most widely used spatiotemporal fusion methods
employ the spatial and temporal adaptive reflectance fusion
model (STARFM) and its enhanced version (ESTARFM)
[43], [94]. Although these were originally proposed for surface
reflectance, they can be applied to other environmental
variables, such as the normalized difference vegetation index,
leaf area index, and LSTs. Over the past decade, various
spatiotemporal fusion methods have been proposed
for obtaining LSTs with high spatiotemporal resolution,
including 1) weighted function-based, 2) unmixing-based,
3) hybrid, and (4) learning-based methods. An example of
weighted function-based spatiotemporal fusion methods is
shown in Figure 9.
MODIS LST
Landsat LST
FIGURE 9. The weighted function-based spatiotemporal fusion methods. The two pairs of
MODIS and Landsat LST data (time t1 and t2) and one MODIS LST data set at the predicted
time t0 can be used to generate Landsat-like LST data of the predicted time t0.
SEPTEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
123
IEEE Geoscience and Remote Sensing Magazine - September 2021

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