Land Cover Detection with Unsupervised Clustering and Hierarchical Partitioning

The classification of a satellite image into land cover classes can be addressed at the level of pixels or segments generated by a image segmentation technique. The segmentation of an image into meaningful regions can be achieved with numerous methods. Recently, a segmentation technique relying on spatial clustering related to the single linkage approach has been put forward [1]. This technique leads to a unique partition of the image domains into maximal segments satisfying a series of constraints related to local (®) and global (!) intensity variations thresholds. It is also suitable for the generation of a hierarchy of partitions by varying the threshold values. However, the generated segments are not classified into thematic classes such as land cover classes. The aim of this work was to apply data clustering to the generated segments in order to i) measure the influence of segmentation on clustering separability; ii) evaluate which threshold values for segmentation maximise the cluster separability; and iii) assess if the clustering could be helpful to detect segments corresponding to transition regions between adjacent segments.

POGGIO Laura;  SOILLE Pierre; 

2009-08-10

Dresden University of Technology

JRC53265

http://www.ifcs2009.de/,    https://publications.jrc.ec.europa.eu/repository/handle/JRC53265,