Full metadata record
DC FieldValueLanguage
dc.contributor.authorFORZIERI GIOVANNIen_GB
dc.contributor.authorMOSER Gabrieleen_GB
dc.contributor.authorCATANI Filippoen_GB
dc.identifier.citationISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING vol. 74 no. 0 p. 175-184en_GB
dc.description.abstractThe potential and limitations of the hyperspectral remote sensing MIVIS sensor (Multispectral Infrared Visible Imaging Spectrometer) in classifying heterogeneous landscapes are explored in this study. In order to quantify the discriminant information derived from selected MIVIS subsets we classified a monitored scenario by progressively increasing the feature space dimensionality. The hyperspectral subsets are defined through the Sequential Forward Selection algorithm, while mapping processes have been performed through the Maximum Likelihood, Spectral Angle Mapper and Spectral Information Divergence classifiers. Impacts of spectral bands on the overall classification accuracies and single land cover-scale reliability, as well as possible dimensionality effects (Hughes phenomenon) are investigated. The analysis is tested on a 20-km stretch of the Marecchia River (Emilia Romagna, Italy) by using MIVIS data acquired in autumn 2009 and 2010 for a 17-class mapping including complex urban/rural areas. For the considered dataset, the MIVIS sensor showed an equipment failure: of the nominal 102-band MIVIS dataset, only the first 24 bands, spanning within the 0.441–1.319 μm spectral range, were exploitable. Nevertheless, the available information provided valuable discriminant contributions in land cover mapping (Maximum Likelihood Overall Accuracy ∼85%) with encouraging reliability on mixed forests, croplands, and no-vegetated floodplain patterns, whereas riparian vegetation and urban zones exhibited low classification accuracies. The relationship between the spectral space dimensionality and the minimum training-set size that is necessary to achieve a given inter-class separability has also been experimentally investigated by progressively under-sampling the original training set. The maximum under-sampling factor that avoided a decrease in the overall accuracy turned out to be, at maximum, 15 for the considered data set.en_GB
dc.description.sponsorshipJRC.H.7-Climate Risk Managementen_GB
dc.titleAssessment of hyperspectral MIVIS sensor capability for heterogeneous landscape classificationen_GB
dc.typeArticles in periodicals and booksen_GB
JRC Directorate:Sustainable Resources

Files in This Item:
There are no files associated with this item.

Items in repository are protected by copyright, with all rights reserved, unless otherwise indicated.