The impact of selection of reference samples and DEM on the accuracy of land cover classification based on Sentinel-2 data

Abstract

Up-to-date and reliable information on land cover and land use status is important in many aspects of human activities. Knowledge about the reference dataset, its coverage, nomenclature, thematic and geometric accuracy, spatial resolution is crucial for appropriate selection of reference samples used in the classification process. In this study, we examined the impact of the selection and pre-processing of reference samples for the classification accuracy. The classification based on Random Forest algorithm was performed using firstly the automatically selected reference samples derived directly from the national databases, and secondly using the pre-processed and verified reference samples. The verification procedures involved the iterative analysis of histogram of spectral features derived from the Sentinel-2 data for individual land cover classes. The verification of the reference samples improved the accuracy of delineation of all land cover classes. The highest improvement was achieved for the woodland broadleaved and non- and sparce vegetation classes, with the overall accuracy increasing from 51% to 73%, and from 33% to 74%, respectively. The second objective of this study was to derive the best possible land cover classification over the mountain area in Norway, therefore we examined whether the use of the Digital Elevation Model (DEM) can improve the classification results. Classifications were carried out based on Sentinel-2 data and a combination of Sentinel-2 and DEM. Using the DEM the accuracy for nine out of ten land cover classes was improved. The highest improvement was achieved for classes located at higher altitudes: low vegetation and non- and sparse vegetation.