IEEE Geoscience and Remote Sensing Magazine - September 2021 - 119

phenology. In contrast, abrupt changes, such as newly constructed
buildings and sudden natural disasters, are difficult
to reconstruct [42].
SPATIOTEMPORAL RECONSTRUCTION METHODS
Spatiotemporal approaches appear to be the most appropriate
methods for reconstructing degraded remote sensing
data with high variability in both space and time [55], [73].
The most basic strategy is to sequentially apply a spatial and
temporal reconstruction method in which a multitemporal
method is used first; a spatial reconstruction method is then
applied if the first step fails to recover all missing pixels or
identify abnormal pixels [56].
For example, Liu et al. introduced a spatiotemporal reconstruction
framework for recovering missing pixels of FY-2F
LST products that included multitemporal and spatial reconstruction
methods [8]. Duan et al. proposed a spatiotemporal
interpolation module for constructing missing AMSR-E
temperature data caused by orbital gaps between satellite
overpasses [24]. For LST image time series, Weiss et al. proposed
a gap-filling method using neighboring valid data and
data from other times, i.e., different calendar dates or multiannual
data sets [74]. Pham et al. reported good results in
applying a 3D gap-filling method to nine years of LST data
observed over Australia between 2002 and 2011 [75].
Recently, Wang et al. proposed a new method for estimating
MODIS or Landsat LST under cloudy skies that used
solar-cloud-satellite geometry (SCSG) [76]. Their scheme
required multitemporal or synthetic reference LST data (e.g.,
MOD11A2 averaged eight-day LST products) and a neighboring
clear-sky region to calculate a ratio based on the SCSG.
Moreover, some hybrid methods that combine spatial or
temporal approaches with spatiotemporal techniques have
been proposed. These hybrid methods utilize the advantages
of the individual approaches but avoid their disadvantages
[77]. Many studies have demonstrated that hybrid and spatiotemporal
methods usually yield lower errors [78].
A COMPARISON OF METHODS FOR
RECONSTRUCTING MISSING PIXELS
Figure 6 shows reconstruction results for simulated missing
pixels using different reconstruction methods. The original
image in Figure 6(a) is a Terra MODIS LST image acquired on
10 October 2010, and Figure 6(b) is an LST image that simulates
the missing pixels from Figure 6(a). Figure 6(c)-(e)
shows the Terra MODIS LST images acquired on 4 September
2010, 20 September 2010, and 6 October 2010, respectively,
which were used to reconstruct the missing pixels from Figure
6(b). The result of the spatial reconstruction method,
shown in Figure 6(f), includes severe artifacts, particularly in
1 January 2011
Classification Map
1 January 2011
AMSR-E LST
2 January 2011
2,800
2,600
2,400
2,200
2,000
2,800
2,600
2,400
2,200
AMSR-E LST
2,000
2,000
2,200
2,400
2,600
2,800
2,600
2,400
2,200
2,000
2,000
2,200
2,400
2,600
2,800
2,000
2,800
2,600
2,400
2,200
2,000
2,800
2,000
2,200
2,400 2,600
2,800
2,200
2,400 2,600
2,800
2,800
2,600
2,400
2,200
2,000
2,800
2,600
2,400
2,200
2,000
2,000
2,200
2,400
2,600 2,800
Reconstruction
2,000
2,200
2,400
2,600 2,800
AMSR-E LST
FIGURE 5. The spatiotemporal reconstruction of missing AMSR-E LST pixels. The target (observed on 1 January 2011) and multitemporal
AMSR-E LST images (observed on 2 January 2011) are used to reconstruct the missing LST pixels for generating a spatially continuous
AMSR-E LST (1 January 2011) [56]. The scatterplots between the target and multitemporal AMSR-E LST images correspond to those on the
classification map.
SEPTEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
119

IEEE Geoscience and Remote Sensing Magazine - September 2021

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