Sentinel-2 derived central Lena Delta land cover classification

• Main Authors: Landgraf, Nele, Shevtsova, Iuliia, Pflug, Bringfried, Heim, Birgit
• Format: Dataset
• Language: English
• Published: PANGAEA
• Subjects: AWI_Envi; Carbon in Permafrost / Kohlenstoff im Permafrost; KoPF; Land cover; Lena Delta; Polar Terrestrial Environmental Systems @ AWI; Sentinel-2; Arctic; lena delta; permafrost
• Online Access: https://doi.pangaea.de/10.1594/PANGAEA.945057

This data set collection represents a Land Cover classification derived from one high quality cloudless optical Sentinel-2A (S2) acquisition, 6. August 2018. Source: atmospherically corrected Sentinel-2 surface reflectance (DLR Adlershof Sentinel-2 processing) The data set collection contains i) 1-band raster file in geotiff format with assigned land cover classes CRS: EPSG:32652 - WGS 84 / UTM zone 52N, 2201 rows, 2930 columns Band 1 is the Land Cover class band, with values from 0 to 11 (12 classes) Classifyer: Random Forest, trained with ROIs (LD18 vegetation plots + manually labelling using expert knowledge) ii) 10 training classes representing different vegetation composition in the form of ESRI polygon shape files, covering Lena Delta 2018 expedition (LD18) vegetation plots and extended with expert knowledge from the field (file names are indicative) iii) 92 training elements representing different vegetation composition in the form of Elementary Sampling Units ESUs: 23 true ESUs representing the LD18 vegetation plots and 69 pseudo ESUs set with expert knowledge from the field and from Lena Delta expedition field reports. the source of the 26 LD18 vegetation plots is Shevtsova et al (2021): Foliage projective cover of 26 vegetation sites of central Lena Delta from 2018. PANGAEA, https://doi.pangaea.de/10.1594/PANGAEA.935875 We atmospherically corrected all resampled all S-2 spectral bands and resampled to 10 m pixel resolution- We calculated and added one NDVI band (NIR-RED / NIR + RED) to the input band collection. We tested several classifiers, also with different selected band combinations. The classification was forced to express vegetation and moisture regimes. We took out the water classes (transparent to turbid) and the sandbank surfaces from the classification by masking them inactive using a threshold method (water mask is based on S-2 NIR 10 m band, sand mask is based on the S-2 BLUE 10 m band). Best results were obtained by using the random forest classification with a band combination of all ...