ON THE EFFECTIVENESS OF SENTINEL-2 DATA FOR LAND-COVER MAPPING: COMPARISON WITH LANDSAT AND SPOT IMAGERY

摘要

The objective of this work is twofold. On the onernhand, we aim at assessing the effectiveness ofrnSentinel-2 data for land-cover mapping, andrnevaluating the improved discrimination capabilitiesrnoffered by new features of the Multi-SpectralrnImager (MSI) sensor. On the other hand, werncompare the performances with those obtainedrnusing both Landsat-5 TM and SPOT-5 HRGrnimagery.rnSimulated Sentinel-2 data are derived fromrnhyperspectral airborne images acquired in thernframework of four different ESA campaigns,rnnamely SPARC 2003 (Barrax, Castilla-La Mancha,rnSpain), AGRISAR 2006 (Demmin, Pomerania,rnGermany) and CEFLES2 2007 (Marmande,rnAquitaine, France). In each case, we discard thernthree spectral bands at 60 meter resolution (i.e.,rnband 1, band 9 and band 10) and resample all the 20rnmeter-resolution bands to 10 meter resolution usingrnnearest neighbour interpolation. Available priorrnknowledge is used for defining a complete groundrntruth for all the land-cover classes characterizingrneach investigated site.rnIn each case, besides considering the wholernavailable 10 spectral bands, we also run the branchrn& bound feature selection algorithm for identifyingrnthe subset of n features (varying n from 1 to 9)rnmaximizing the (expected) separability between therninvestigated land-cover classes (for which trainingrnsamples are available). Furthermore we runrnexperiments by adding the new features of Sentinel-rn2 successive to the corresponding Landsat-5rnThematic Mapper (TM) bands. Then, in order tornassess the discrimination capabilities offered byrndifferent features, for each subset we run twornsupervised classifiers, namely, the MaximumrnLikelihood (ML) classifier and Support VectorrnMachines (SVM). ML is a simple yet generallyrnrather effective statistical classifier, which does notrnrequire the user to set any free parameter. SVM arernadvanced state-of-art classifiers, which provedrncapable of outperforming other traditionalrnapproaches. For the selection of the two freernparameters (i.e., a penalization parameter and thernvariance of considered Gaussian kernels) wernemployed a 5-fold cross-validation strategy.rnFor each case study we also simulate correspondingrnLandsat-5 Thematic Mapper (TM) and SPOT-5rnHigh Resolution Geometric (HRG). Then, wernresample simulated images to 10 meters spatialrnresolution (using nearest neighbour interpolation)rnand run both ML and SVM classifiers.rnPerformances are finally compared to thosernobtained with Sentinel-2 data and evaluated bothrnqualitatively and quantitatively (in terms ofrnpercentage of overall accuracy and kapparncoefficient of accuracy).rnPromising results obtained so far confirm therneffectiveness of Sentinel-2 data for land-coverrnmapping as well as the improved discriminationrncapabilities with respect to Landsat-5 TM andrnSPOT-5 HRG data.