IEEE Geoscience and Remote Sensing Magazine - September 2021 - 124

WEIGHTED FUNCTION-BASED METHODS
Since they were first proposed, weighted function-based
methods, such as STARFM and ESTARFM, hereafter referred
to collectively as STARFM-like methods, have gained considerable
attention. Some studies directly adopted STARFMlike
methods to generate daily fine-resolution thermal data
and their derivative products, such as evapotranspiration
and LST [95]-[99]. However, their performance, to a certain
degree, degrades over heterogeneous areas because obvious
differences exist between reflectance and LST.
STARFM-like methods with various improvements
were designed for generating high-spatiotemporal-resolution
LSTs. These enhancements were made in the model
relationships among different-resolution pixels, weighted
function design, number of sensors, and the temporal
variation modeling or thermal landscape representation,
which aimed to improve spatiotemporal LST patterns.
With the aid of the weighted function, linear relationships
were established between two MODIS LSTs of different
times or a MODIS and Landsat LST of the same time to predict
synthetic Landsat-scale LST data [100], [101].
Based on the STARFM framework, Huang et al. and Wu
et al. designed bilateral filtering- and variation-based fusion
methods, respectively, for producing high-spatiotemporalresolution
LSTs with higher accuracy [102], [103]. Weng et al.
and Quan et al. modified ESTARFM to obtain LSTs with high
spatiotemporal resolution by considering the annual temperature
cycle (ATC) and urban thermal landscape heterogeneity
[104], [105], respectively. Wu et al. proposed a spatiotemporal
integrated temperature fusion model (STITFM)
for generating fine-spatiotemporal-resolution LSTs with
more than two sensors [17]. Weighted function-based
methods generally have shown good performances in
the literature and can be improved further by considering
complex surface heterogeneity and rapid changes in the
surface type.
UNMIXING-BASED METHODS
Unmixing-based methods regard temporal variation at a
low-spatial-resolution pixel as a mixture of some component
temporal variations at the corresponding high-spatial-resolution
pixels. As a pioneering work, an unmixingbased
fusion idea to combine multisensor satellite images
observed at different times with different resolution was
proposed [106].
Generally speaking, unmixing-based methods have two
main challenges: serious errors associated with the unmixing
process and a lack of within-class variability of the finescale
information inside a coarser pixel [46]. Subsequently,
unmixing-based approaches were designed to address these
challenges. For example, specific constraints were incorporated
into a unmixing process to ensure that the obtained reflectance
values were positive and within a physically meaningful
range [107]. To account for land cover change, a spatial
and temporal data fusion approach (STDFA) and its modified
version (ESTDFA) were put forward that use two or more
124
image pairs with a constant and an adaptive-moving window,
respectively [108], [109].
Huang and Zhang described an unmixing-based fusion
technique capable of considering phenological differences
and changes in the surface type [110]. However, owing to
the complexity of temperature unmixing, these methods
are often applied to reflectance data but not to LST data.
Wu et al. estimated LSTs with high spatiotemporal resolution
for environmental-process monitoring using STDFA
directly and compared the results with those obtained by
STARFM and ESTARFM [111]. Their results indicated that,
although all three methods captured the spatiotemporal
change information of the LSTs, they also showed performance
differences, with ESTARFM, STDFA, and STARFM
performing best to worst, respectively. In summary, unmixing-based
methods are widely limited by inaccurate
estimation numbers of endmembers, endmember spectral
variability in multitemporal data, and spectral mixing nonlinearities.
Furthermore, these techniques are challenged
by land cover changes.
HYBRID METHODS
Hybrid methods usually integrate two or more techniques
and contain weighted function- or unmixing-based approaches.
Their purpose is to enhance the performance of
data fusion by integrating the advantages of different methods.
Representative approaches include the spatial and temporal
reflectance unmixing model [112], flexible spatiotemporal
data fusion (FSDAF) [113], the spatial-temporal
remotely sensed images and land cover maps fusion model
[114], bleed spatiotemporal temperature (BLEST) [105], and
weighted combination of kernel-driven and fusion-based
methods (CKFM) [50].
The first three methods combine ideas from unmixing
theory, weighted function, Bayesian theory, and spatial interpolation.
Hybrid methods can fuse reflectance images
under complex conditions, e.g., heterogeneous landscapes
or abrupt land cover changes [46]. Inspired by the spatiotemporal
adaptive data fusion algorithm for temperature
mapping [104] and FSDAF, the BLEST method combines
multiple strategies for predicting hourly Landsat-like LSTs
in heterogeneous areas. Specially, BLEST adopts weighted
functions to preserve more details in LST images of
fine spatial resolution. Moreover, it uses ATC and DTC
models to describe nonlinear temporal patterns of LST
variations between LST images of coarse resolution, and
it combines the temperature-mixing model and downscaling
technology to account for effects from complex
and changeable landscapes.
However, assuming a linear or simple nonlinear relationship
between the input and predicted LST data over
complex landscapes might not be a sufficiently accurate
approximation of reality. Furthermore, the literature
shows that kernel-driven methods can capture abundant
detail of the TIR band from visible bands of finer
spatial resolution, whereas fusion-based techniques are
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
