IEEE Geoscience and Remote Sensing Magazine - June 2017 - 51

■ Asphalt
■ Meadows
■ Gravel
■ Trees
■ Painted

Metal Sheets

■ Bare Soil
■ Bitumen
■ Self-Blocking

Bricks

■ Shadows

Test Samples

Reference (OA: 72.6%) HS (OA: 62.4%)
(a)

MS (OA: 64.6%)

HySure (OA: 83.7%)

■ Healthy Grass
■ Stressed Grass
■ Trees
■ Soil
■ Water
■ Residential
■ Commercial
■ Road
■ Highway
■ Parking Lot 1
■ Parking Lot 2

Test Samples

Reference (OA: 74.8%) HS (OA: 67.0%)
(b)

MS (OA: 69.9%)

HySure (OA: 78.5%)

FIGURE 18. The classification maps of the reference, HS, MS, and HySure images obtained by RoFs for the (a) University of Pavia and (b)
University of Houston data sets.

results, whereas ECCV'14 resulted in poor classification
accuracies for both data sets.
3) Noticeably, for some methods, the relative classification performance does not confirm the ranking in
the statistical quantitative assessment tests discussed
earlier. GSA, for instance, showed a high potential
for classification purposes even though it was not
competitive in the previously conducted quantitative
assessment tests.
4) Some of the fused images led to higher classification
accuracies than those of the reference images, which is
mainly due to the denoising effect of HS-MS fusion or
denoising postprocessing. Higher SNRs of the fused images can mitigate salt-and-pepper misclassified pixels.
It should be noted that classification-based validation of HS-MS fusion results does not reflect the absolute
quality of fused data-firstly, because the ground-truth
information is available only for a comparatively small
spatial subset of the data and, secondly, because of the
many factors and steps in the data preprocessing chain
that may influence the classification capabilities relative to
the ground-truth data. The characteristics of applications
and data analysis methods highly influence the validation results. It is possible, for instance, to design a fusion
method specifically for one classification, application, or
data analysis scenario for which it incorporates as much
june 2017

ieee Geoscience and remote sensing magazine

information in its underlying mathematical fusion model
as available. Despite its presumably outstanding performance in that scenario, it would certainly fail in a more
diverse and comprehensive testing setup, such as the one
presented in this article. If one is interested in finding a fusion method whose high-resolution data output is particularly suitable for SVM- or RoF-based classification, the performance comparison presented in this section provides a
good starting point for a more in-depth study tailored to
the scenario of interest.
CONCLUSIONS AND FUTURE RESEARCH
In this article, we presented a comparative study of
HS-MS fusion methods for the resolution enhancement of HS imagery. Ten state-of-the-art HS-MS fusion
methods based on four different approaches (CS, MR A,
unmixing, and Bayesian) were analyzed, assessed, and
compared using visual, quantitative, and classificationbased means of evaluation. Publicly available program
codes were used while performing careful parameter
tuning at optimal individual algorithm settings. To
evaluate the generalizability and versatility of HS-MS
fusion performance, eight simulated data sets were used
in the experiments based on various representative fusion scenarios. The overall fusion characteristics of all
the methods under comparison regarding performance
51
Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2017
