IEEE Geoscience and Remote Sensing Magazine - September 2015 - 97

FIgUre 4. Classification result for Houston University using 95% downsampled ET, segmentation with MRF, and man-made structure correction.

Running Track
Tennis Court
Parking Lot 2
Parking Lot 1
Railway
Highway
Road
Commercial
Residential
Water Body
Soil
Tree
SynthetGrass
StressedGrass
HealthyGrass
Non Class
september 2015

ieee Geoscience and remote sensing magazine

fragile to outliers since it repeatedly tries to find a way to
classify them (eventually overfitting them to some extent).
For the two other algorithms, the RF uses randomized sampling for each tree in conjunction with randomized feature
selection for the node splits, which can result in better control of numerous outliers, compared to the fully randomized cut-points selection of ET, that seems better suited to
a smaller number of outliers since it only operates on the
features and not on the training samples (therefore the outliers will always appear in all trees).
In the third experiment, where zero-mean White
Gaussian Noise with different variances is added to the
entire feature space, the performance of the three algorithms is comparable across the different noise levels, with
a small advantage for the ET algorithm. It is interesting to
note that, while GBRT is typically the least performing of
the three when considering only the classification part, it
is again typically improving more than the other two algorithms from postprocessing, for the reasons that were
mentioned previously. It should be noted that this effect is
very clear for the particular dataset, since post-processing
segmentation is the only step applied here, while for the
University of Houston the human structure correction step
that is done in addition is favoring algorithms that have a
high initial performance for the majority of the pixels in
each of the extracted structures and this may balance out
the described effect.
Another interesting observation is that in the Gaussian
noise setting the good complementarity between GBRT and
MRF post-processing can also be observed from the standard deviation of the final result, which is comparable and
in some cases less than the respective standard deviation
of the final result produced by the other two methods. On
the other hand, in the previous setting where outliers were
injected, an erroneous classification result from GBRT directly translates to a bad final (post-processed) outcome.
Therefore, in the cases that the outliers distorted classification the final outcome was significantly worse compared to
cases that the outliers had less impact, thus resulting also in
higher standard deviation for GBRT compared to the other
two methods.
6. CONCLUSIONS
In this paper we focused on the use of ensemble methods
in remote sensing. We briefly described relevant state of
the art in ensemble learning, both inside and outside the
remote sensing community and elaborated on the theoretical basis of this research field. Following this introduction,
we provided a detailed review of three of the most relevant
and prominent techniques in ensemble learning, namely
the Random Forest, Extra Trees and the Gradient Boosted
Regression Trees algorithms. All algorithms were assessed
in terms of their theoretical properties as well as applicability for remote sensing use cases. Special care was to taken
to describe the effect of hyper-parameters and other configuration choices in a way that is useful for the practitioner
97
Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2015
