Boosting vs maximum de vraisemblance
Arnaud, Michel | Bailly, Jean-Stéphane | Puech, Christian | Centre de Coopération Internationale en Recherche Agronomique pour le Développement (Cirad) | AgroParisTech | Structures et Systèmes Spatiaux (UMR 3S) ; Ecole Nationale du Génie Rural, des Eaux et des Forêts (ENGREF)-Centre national du machinisme agricole, du génie rural, des eaux et forêts (CEMAGREF)
This article sets out to demonstrate how boostingcan serve as a supervised classification method and to compare its results with those of conventional methods. The comparison begins with a theoretical example drown from remote sensing classification in which several criteria are varied: number of pixels par class, overlapping (or not) of radiomatric values between classes, with and without spatial structuring of classes within the geographical space. It is seen that 1) maximum likelihood give better results than boosting when the radiometric values for each class are clearly separated. This advantage is lost as the number of pixels par class increases; 2) boosting outperforms maximum likelihood in the event of overlapping radiometric variable classes, wether or not there is a spatial structure.
Показать больше [+] Меньше [-]Ключевые слова АГРОВОК
Библиографическая информация
Эту запись предоставил AgroParisTech