Selection criteria of training set for optimal land cover discrimination in automatic segmentation

This work is part of a wider project whose general objective is to develop an automatic classification methodology, congruent with the CORINE land-cover thematic legend, by high resolution multi-spectral IKONOS images datasets. The imagery were provided by Regione Marche institution thanks to a research agreement signed together with the Technical University of Marche jointed departments (DARDUS, DIIGA and DiSASC). According to a hierarchical approach, different phases exist for image classification. At the lowest level classification was merely based on specific pixels spectral value, while moving toward highest levels, segmentation methodology can be used according to features spatial pattern to perform better results. The present paper aims to deepen the first of the above mentioned hierarchical levels. The methodological approach focus on the optimal selection criteria, in order to define the best setting (spectral bands, Ground Truth, etc.) for the training stage. With this intent different supervised classification algorithms have been tested. In particular, the training stage was carried out tacking advantage of a dedicated-GIS platform implementation. Each Ground Truth sample was collected by means of specific campaign and/or pan-sharpened IKONOS dataset visual interpretation. The depicted classes were grouped in different levels of increasing detail. To be more confident, the CORINE standard legend has been modified according to current study case specificity, in order to obtain an optimal distribution of the samples in the training set. This also allowed to improve the training stage setting by excluding clusters whose spectral values largely ranging far from corresponding average class values.