Endmember Extraction Filter

Extracts endmembers from an input training data set.

Category	Decomposition
Node
Parameters	Composition: which algorithm is used to determine the endmembers: Group Average: the averages of the training data assigned to the applicable groups will be considered to be the endmembers (this doesn’t actually calculate anything, it’s main usage is to provide an easy way to use group averages as the input of an Abundance Determination Filter) Pure Pixel: the filter will pick input data points as endmembers that best can be used to span a simplex encompassing the input (but each endmember is a spectrum that actually occurs in the input data) Mixed Pixel: the filter will find the endmembers as points that span a simplex that encompass the entire input data (the endmembers are not themselves actual spectra that appear in the input) Determination: the starting conditions for the endmember extraction algorithm (only relevant if Composition is not Group Average): Unsupervised: groups will be ignored and all endmembers will be extracted automatically from the entire set of input training data Group Supervised: the endmembers will be associated with the applicable groups (hence the groups influence which endmembers will be chosen) Endmembers: the number of endmembers to extract in the Pure Pixel case. If Determination is Group Supervised this gives the number of endmembers per group. Iterations (only relevant in the Mixed Pixel case): the maximum number of iterations to perform the extraction algorithm for Refinement (only relevant in the Pure Pixel case): whether to perform an additional refinement step on the endmembers found Applicable (per group): whether a group is applicable to this filter, only relevant for the Group Average and Group Supervised cases.
Inputs	Input: the input data
Outputs	Endmembers: the endmembers that were determined from the input data (This is training only data that will only be used indirectly during execution)

Effect of the Filter

The filter is generally used in combination with the Abundance Determination Filter. It calculates endmembers, i.e. spectra that are representative for the entire training data set. Endmembers are generally chosen so that they can be used to span a simplex encompassing the training data. Each spectrum within the training data may then be represented as a combination of the endmembers.

The output of the filter consists of training-only data. Its only purpose is to be input to other trained filter, such as the aforementioned Abundance Determination Filter.