Request for a native GRASS implementation of the ISODATA clustering algorithm

[Nikos Alexandris]

The ISODATA clustering algorithm

The ISODATA algorithm is a(nother) modification of the k-means algorithm. Roughly, the difference lies in that the former allows for user-defined number of clusters while the latter pre-assumes a fixed number of clusters.

Quoting [details from A Fast Implementation of the ISODATA Clustering Algorithm][1]:

...the algorithm tries to find the best set of cluster centers for a given set of points in d-dimensional space through an iterative approach until some maximum number of iterations are performed. It uses a number of different heuristics to determine when to merge or split clusters. [[BR]]
 
At a high level, in each iteration of the algorithm the following takes place: points are assigned to their closest cluster centers, cluster centers are updated to be the centroid of their associated points, clusters with very few points are deleted, large clusters satisfying some heuristics are split, and small clusters satisfying other heuristics are merged. The algorithm continues until maximum number of iterations are performed. Here we go over the algorithm in more detail. See the related publications below for further information. [[BR]]

Clustering in GRASS

The current clustering implementation in GRASS GIS' module ​i.cluster is also another modification of the k-means clustering algorith. One difference, from an end-user point-of-view, is that i.cluster exptects at least two input raster maps (variables) in order to run.

Differences for end-users

The ISODATA algorithm, contrary to i.cluster, can run on a single raster map (variable) which makes it an easy and quick alternative for pixel-based unsupervised image classification tasks.

This makes it, be it a good or bad choice, attractive for many (commercial) remote sensing projects.

Proprietary Implementations

Most proprietary & commercial GIS & Remote Sensing packages have integrated an option for the ISODATA unsupervised (and pixel-based) classification algorithm. Some sources below:

• Iso Cluster Unsupervised Classification (Spatial Analyst) [2] in ArcGIS
  
• ISODATA Clustering [3] in ERDAS
  
• Classification Tutorial (PDF) [4] for ENVI
  
• Technical Specifications, Geomatica Core (PDF [5] about Geomatica
  
• Idrisi Selva Technical Specifications (PDF) [6] about Idrisi Selva
  

FOSS Implementations

Some FOSS implementations of the ISODATA clustering/classification algorithm include:

• the "Fast Implementation of the ISODATA Clustering Algorithm" by Nargess Memarsadeghi, David M. Mount, Nathan S. Netanyahu, and Jacqueline Le Moigne [1]
  which has found its way in ​Opticks [7]
• the biOps package for R [8]
  

However, those are not easy alternatives for a GRASS GIS user.

References

[1] ​http://www.cs.umd.edu/~mount/Projects/ISODATA/
[2] ​http://help.arcgis.com/en/arcgisdesktop/10.0/help/index.html#//009z000000pn000000.htm
[3] ​http://geospatial.intergraph.com/fieldguide/wwhelp/wwhimpl/js/html/wwhelp.htm
[4] ​http://www.exelisvis.com/portals/0/tutorials/envi/ClassificationTutorial.pdf
[5] ​http://www.pcigeomatics.com/products/pdfs/Geomatica_Core_1032.pdf
[6] ​http://www.clarklabs.org/products/upload/IDRISI-Selva-GIS-Image-Processing-Specifications.pdf
[7] ​http://cran.r-project.org/web/packages/biOps/index.html
[8] ​http://www.opticks.org

[benducke]

GPL'd implementations in C++ can be found here: ​http://www.cs.umd.edu/~mount/Projects/ISODATA/

[Nikos Alexandris]

Link to "historical" i.cluster material (kindly pointed to by Markus N)

​http://lists.osgeo.org/pipermail/grass-user/1997-December/000912.html

[martinl]

Milestone renamed

[neteler]

Ticket retargeted after milestone closed

[martinl]

All enhancement tickets should be assigned to 7.6 milestone.

[martinl]

Ticket retargeted after milestone closed

[martinl]

Ticket retargeted after milestone closed