Morpho-colorimetric and non-parametric analysis in statistical classification of vascular flora

This dissertation concerns the task of classifying objects by methods of morpho-colorimetric and non-parametric statistical analysis. The study can be applied in botany where manual classification of seeds is still a common practice. It is labor-intensive, subjective, and suffers from contradiction, as well as it is a time-consuming task even for highly specialized botanists. Starting from this problem, automated, consistent, and efficient algorithms of classification of seeds have been developed allowing the researcher to have a valid support for reducing drastically time for classification and, at the same time, an exploratory tool that detects latent patterns in data that human eye cannot identify. Firstly an approach called Background Subtraction, that enhances the quality of segmentation process output of an image, has been proposed. From RGB images it allows to get more precise binary images which need of a reduced intervention of manual correction. Then for each object data concern size, texture, color and shape have been extracted. These have been used as input for classification process, in which four classifiers have been performed: Linear Discriminant Analysis (LDA), Classification And Regression Trees (CART), Support Vector Machines (SVM) and Naïve Bayes (NB). In order to enhance the classification accuracy an approach of classifier combination, indicated as CA, has been developed. It consists in spliting the complex problem of classifying among D classes into D−1 sub problems less complex than the original one, each of them classifying between only two classes. Combining the four classifiers considered, CA allowed to reduce of 25% the misclassification error obtained by the best of the four classifiers. Finally, approach aimed at evaluating the reliability of a classification rule has been proposed. The algorithms proposed are developed and optimized for botanical seeds, but they are suitable to a larger class of morphological classification problems. In order to make these algorithms usable and executable, functions have been created in R language and published.