Centre de documentation
scientifique

Semi-supervised SVM for individual tree crown species classification / Michele Dalponte in ISPRS Journal of photogrammetry and remote sensing, vol 110 (December 2015)  

Public [article]  inISPRS Journal of photogrammetry and remote sensing > vol 110 (December 2015) . - pp 77 – 87 Titre : Semi-supervised SVM for individual tree crown species classification Type de document : Article/Communication Auteurs : Michele Dalponte, Auteur ; Levi Theodor Ene, Auteur ; Mattia Marconcini, Auteur ; et al., Auteur Année de publication : 2015 Article en page(s) : pp 77 – 87 Note générale : bibliographie Langues : Anglais (eng) Descripteur : [Vedettes matières IGN] Traitement d'image optique [Termes IGN] classification dirigée [Termes IGN] classification par séparateurs à vaste marge [Termes IGN] données laser [Termes IGN] forêt boréale [Termes IGN] image hyperspectrale [Termes IGN] inventaire forestier local Résumé : (auteur) In this paper a novel semi-supervised SVM classifier is presented, specifically developed for tree species classification at individual tree crown (ITC) level. In ITC tree species classification, all the pixels belonging to an ITC should have the same label. This assumption is used in the learning of the proposed semi-supervised SVM classifier (ITC-S3VM). This method exploits the information contained in the unlabeled ITC samples in order to improve the classification accuracy of a standard SVM. The ITC-S3VM method can be easily implemented using freely available software libraries. The datasets used in this study include hyperspectral imagery and laser scanning data acquired over two boreal forest areas characterized by the presence of three information classes (Pine, Spruce, and Broadleaves). The experimental results quantify the effectiveness of the proposed approach, which provides classification accuracies significantly higher (from 2% to above 27%) than those obtained by the standard supervised SVM and by a state-of-the-art semi-supervised SVM (S3VM). Particularly, by reducing the number of training samples (i.e. from 100% to 25%, and from 100% to 5% for the two datasets, respectively) the proposed method still exhibits results comparable to the ones of a supervised SVM trained with the full available training set. This property of the method makes it particularly suitable for practical forest inventory applications in which collection of in situ information can be very expensive both in terms of cost and time. Numéro de notice : A2015-894 Affiliation des auteurs : non IGN Thématique : IMAGERIE Nature : Article nature-HAL : ArtAvecCL-RevueIntern DOI : 10.1016/j.isprsjprs.2015.10.010 En ligne : http://dx.doi.org/10.1016/j.isprsjprs.2015.10.010 Format de la ressource électronique : URL article Permalink : https://documentation.ensg.eu/index.php?lvl=notice_display&id=79445 [article]