Unsupervised Classification & Segmentation
eCognition offers a comprehensive number of geoscience analysis methods: supervised & unsupervised machine learning, deep learning, knowledge-based analysis with Fuzzy Logic or threshold conditions. All together to generate or classify objects.
For unsupervised classification, eCognition users have the possibility to execute a ISODATA cluster analysis. ISODATA stands for “Iterative Self-Organizing Data Analysis Technique” and categorizes continuous pixel data into classes/clusters having similar spectral-radiometric values. To label thematic information to the unknown classes is the task of the user afterwards.
Unsupervised classification is appropriate when the definitions of classes, and perhaps even the number of classes, are not known in advance. That means cluster analysis is meaningful if you do not know much about the data beforehand and want to identify spectral clusters in the layers to learn more about the input data/sensor/region.
The output of a cluster analysis in eCognition is normally a raster layer and customer can use this raster layer to generate and classify objects. But the cluster analysis layer can also be used to improve a thematic classification or to optimize object outlines.
To facilitate unsupervised classification in eCognition, Christian Weise created (in 2014) an customized algorithm for ‘unsupervised classification and segmentation’: Users have just to define the input domain and the algorithm will use all available raster layers of the input domain for a cluster analysis, generate objects for each cluster and classify clusters in the given object level. Very easy to use.
To use the ‘unsupervised classification and segmentation (ISODATA clustering)’ please download the customized algorithm and load it into your project, after it you will find the ‘unsupervised classification and segmentation (ISODATA clustering)’ in the algorithm list. The ‘Example Project - Unsupervised Classification’ provides some use cases (starting with theoretically examples to learn how to use the algorithm).
Was this article helpful?