Centre de documentation
scientifique

Per-pixel bias-variance decomposition of continuous errors in data-driven geospatial modeling : A case study in environmental remote sensing / Jing Gao in ISPRS Journal of photogrammetry and remote sensing, vol 134 (December 2017)  

Public [article]  inISPRS Journal of photogrammetry and remote sensing > vol 134 (December 2017) . - pp 110 - 121 Titre : Per-pixel bias-variance decomposition of continuous errors in data-driven geospatial modeling : A case study in environmental remote sensing Type de document : Article/Communication Auteurs : Jing Gao, Auteur ; James E. Burt, Auteur Année de publication : 2017 Article en page(s) : pp 110 - 121 Note générale : Bibliographie Langues : Anglais (eng) Descripteur : [Vedettes matières IGN] Traitement d'image [Termes IGN] apprentissage automatique [Termes IGN] classification pixellaire [Termes IGN] décomposition [Termes IGN] données environnementales [Termes IGN] erreur absolue [Termes IGN] erreur systématique [Termes IGN] image Landsat [Termes IGN] précision de l'estimation [Termes IGN] surface imperméable [Termes IGN] test de performance [Termes IGN] variance Résumé : (Auteur) This study investigates the usefulness of a per-pixel bias-variance error decomposition (BVD) for understanding and improving spatially-explicit data-driven models of continuous variables in environmental remote sensing (ERS). BVD is a model evaluation method originated from machine learning and have not been examined for ERS applications. Demonstrated with a showcase regression tree model mapping land imperviousness (0–100%) using Landsat images, our results showed that BVD can reveal sources of estimation errors, map how these sources vary across space, reveal the effects of various model characteristics on estimation accuracy, and enable in-depth comparison of different error metrics. Specifically, BVD bias maps can help analysts identify and delineate model spatial non-stationarity; BVD variance maps can indicate potential effects of ensemble methods (e.g. bagging), and inform efficient training sample allocation – training samples should capture the full complexity of the modeled process, and more samples should be allocated to regions with more complex underlying processes rather than regions covering larger areas. Through examining the relationships between model characteristics and their effects on estimation accuracy revealed by BVD for both absolute and squared errors (i.e. error is the absolute or the squared value of the difference between observation and estimate), we found that the two error metrics embody different diagnostic emphases, can lead to different conclusions about the same model, and may suggest different solutions for performance improvement. We emphasize BVD’s strength in revealing the connection between model characteristics and estimation accuracy, as understanding this relationship empowers analysts to effectively steer performance through model adjustments. Numéro de notice : A2017-731 Affiliation des auteurs : non IGN Thématique : IMAGERIE Nature : Article nature-HAL : ArtAvecCL-RevueIntern DOI : 10.1016/j.isprsjprs.2017.11.001 En ligne : https://doi.org/10.1016/j.isprsjprs.2017.11.001 Format de la ressource électronique : URL article Permalink : https://documentation.ensg.eu/index.php?lvl=notice_display&id=88429 [article]

Exemplaires(3)

Code-barres	Cote	Support	Localisation	Section	Disponibilité
081-2017121	RAB	Revue	Centre de documentation	En réserve L003	Disponible
081-2017122	DEP-EAF	Revue	Nancy	Dépôt en unité	Exclu du prêt
081-2017123	DEP-EXM	Revue	Saint-Mandé	Dépôt en unité	Exclu du prêt