Minimum Redundancy Maximum Relevance (mRMR)
Introduction
The minimum-redundancy-maximum-relevance criteria is a
feature selection method which helps in finding the optimal
visualization of an image with respect to a target class. The
features are ranked according to their redundancy and relevance and
the first three are used to obtain the improved representation.
DAS-Tool module applies this algorithm on the desired spectral
bands of the product to identify the most relevant 3 bands used to
display it.
Algorithm
Each pixel is represented by x bands, and a patch
contains M x N pixels. Given the fact
that is no need for external parameters, having the number of
required features, 3, the basic algorithm implemented in SNAP
is:
- Consider each patch being part of a single semantic class and
initialize the reference vector VMxN with elements of
the same value;
- Apply minimum-redundancy-maximum-relevance criteria on bands
values;
- Return the first 3 ranked bands to be used for the improved
image display.
User Interface
The mRMR visualization tool can be invoked in two
ways:
- From the Product Explorer window, right click on the
product to visualize, and select Open mRMR Image Window
- From the Optical menu, select DAS-Tool ->
Exploratory Visual Analysis -> Open nRMR Image Window
If the product's bands have the same sizes, selecting the Open FA
Image Window command, the following dialog box will appear:
otherwise the following warning message will appear:
Preview Panel Group
Scene Preview:
Used to display the grey image of the product to be processed, and
it allows the user to select the desired area (patch) on which the
algorithm will be applied.
Patch Preview:
Used to display the patch on which the algorithm will be applied.
Patch settings group
Start coordinates:
Used to display or set the coordinates of the top-right corner of
the patch on which the algorithm will be applied.
Patch size:
Used to select the size of the patch (square area).
Manual selection:
By selecting this radio button, the user will be able to choose
the area on which the algorithm will be applied. The selection can
be made directly from the
Scene preview
or by setting the coordinates of the top-right corner of the area
from the
Start coordinates
.
Automatic selection:
By selecting this radio button, the area used by the algorithm,
will be automatically selected by the system so that its content is
as diverse as possible.
mRMR computed bands group
Red, Green, Blue:
display the most relevant 3 bands identified by the algorithm and
which will be used to display the image.
References
H. Peng, F. Long, and C. Ding,
Feature selection based on mutual information. Criteria of
Max-Dependency, Max-Relevance and Min-Redundancy, IEEE Trans.
Pattern Anal. Machine Intell., vol. 24, no.8, pp. 1667?1238, Aug.
2005.
D. Bratasanu, I. Nedelcu and M. Datcu,
Interactive Spectral Band Discovery for Exploratory Visual Analysis
of Satellite Images, IEEE Journal of selected topics in applied
Earth Observations and Remote Sensing, vol. 5, no. 1, Feb 2012
