The analysis of images in hyperspectra data, generate sets of prototypes in high dimensional spaces. When the space is analyzed, appear a lot of redundant information, and it is possible to realise a reduction of information without significant loss of information and class discrimination. Several works have studied this empirical question to characterize the number of instances in the training set for a subset of features given.

On the other hand, the increase of features and instances supposes a higher computational cost for the classification associated. One way to perform the classification is stabilizing that decision by means of regularization or injection of noise. Other possibility is the use of ensembles of classifiers as a only classifier rule. There are different methods to generate an ensemble as: Generation of diversity using different training sets (Bagging and Boosting), generation of diversity using subsets of features, and combining different rules of classification.

In this project, a data complexity analysis is realized to predict conditions for prototypes reduction and features reduction in ensembles of classifiers. Finally, the practical context to apply this focus is the study of multispectral images in fruit quality assessment.