IEEE Geoscience and Remote Sensing Magazine - September 2015 - 77

(a)

(b)

Figure 22. Two different scenes (a) and (b).

cloud removal in [45] is as follows: a missing pixel is filled
using an appropriate similar pixel within the remaining regions of the corrupted image, and another reference image
is used as guidance to locate the similar pixels. In order to
ensure the global coherence of the reconstruction image,
a pixel-offset based spatio-temporal MRF (STMRF) global
function is built to jointly select the most suitable similar
pixels in the remaining regions to replace the missing pixels. This STMRF method combines the advantages of the
spatial-based methods, which reconstruct the data using
the radiometric information in the corrupted image itself
to ensure that the recovery keeps a high degree of spectral
coherence and a convincing visual quality, and the temporal-based methods, which guarantee the fine information
accuracy of the reconstruction results.
Fig. 23 illustrates the superior performance of this spatial completion method with temporal guidance. Fig. 23(a)
is the original MODIS image acquired on August 2, 2010,
and Fig. 23(b) is the reference image acquired on March
11, 2010. Fig. 23(c) is the cloud-contaminated image simulated from Fig. 23(a). To allow a comparative analysis, this
hybrid method was compared with a spatial-based method
and a temporal-based method, and their reconstruction
results are shown in Fig. 23(d)-(f). From Fig. 23(d)-(f),
it can be seen that the result of this joint spatio-temporal
method is clearly better than the individual spatial-based
and temporal-based methods. When the terrain is complex
and the missing region is large, the recovery result of the
spatial-based method [Fig. 23(d)] is inaccurate, and the cor-

(a)

(b)

(c)

relation coefficient (CC) calculated from the cloud-contaminated region for this result is just 0.70263. In the result of
the temporal-based method [Fig. 23(e)], most of the ground
features are recovered well, but in the river region, the spectral characteristic is still different from that in the remaining
river region, which may also be a result of the large spectral
differences between the corrupted image and the reference
image. The CC for this result is 0.83804. For the joint spatiotemporal method, it shows the most plausible visual result,
some of the detailed information is recovered well, and the
reconstructed region is the most consistent with the original
image. The CC for this result is the highest, with a value of
0.86527. It is worth noting that the two input images were
acquired in different seasons, and the spectral characteristics of the ground features are significantly different. Most of
the temporal-based methods
cannot effectively deal with
such a difference. However,
the joint spatio-temporal
The TradeoFF beTween
method is better able to adspaTial correlaTion and
dress the issue.
Temporal correlaTion

has an eFFecT on The
B. Joint SpectralreconsTrucTion.
temporal methodS
As is well known, the spectral correlations and temporal correlations are the basic
properties of remote sensing data. When the two correlations are available, it is clearly not enough that only one
of them is utilized. Therefore, the joint spectral-temporal
idea is a promising approach. Currently, the joint spectraltemporal methods [76] are few in number. Here, we introduce one primary example in the framework of sparse representation. Sparse representation has been applied to the
reconstruction of the missing information of remote sensing data, e.g., the aforementioned spectral-based methods
[18], [77], [78] and temporal-based methods [58], [59],
[79]. Their main differences lie in the data organization.
Firstly, it is necessary to review sparse representation. Sparse representation has drawn the attention of a
large number of researchers in recent years. The superiority of the approach is that it has been shown to be able to

(d)

(e)

(f)

Figure 23. Experimental results for the recovery of a cloudy image. (a) Original MODIS image acquired on August 2, 2010. (b) Reference

image acquired on March 11, 2010. (c) Cloud-contaminated image simulated from (a). Recovered images using the following: (d) the spatialbased method in [19]; (e) the temporal-based method in [22]; and (f) the spatial completion method with temporal guidance in [45].
september 2015

ieee Geoscience and remote sensing magazine

77
Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2015
