High-resolution land cover map of High Arctic tundra vegetation in Bjørndalen, Svalbard (resolution: 10 cm)

This dataset provides a 10 cm spatial resolution land cover classification of approximately 140 ha of Arctic tundra in Bjørndalen, Svalbard. The raster classifies the landscape into 14 vegetation and surface cover classes and is distributed as a GeoTIFF together with a color map and boundary files. The land cover classification was created using UAS and in situ vegetation data collected during two field campaigns (11.-18.07.2024, 16.-24.07.2025) at the peak of the vegetation season. The UAS data was acquired using a Trinity Pro (Quantum-Systems GmbH, Gilching, Germany) fixed-wing UAS capable of vertical take-off and landing. The land cover classification map has a spatial extent of (15°17'27.007", 78°12'32.832" : 15°21'0.872", 78°13'31.41") and is provided in EPSG:25833 (ETRS89 / UTM zone 33N). Land cover classification The mapped valley is characterized by a glacial meltwater stream and High Arctic tundra vegetation ranging from sparsely vegetated screes to wetlands. We defined 14 land cover types, differentiated by their dominating species and soil moisture conditions (see README for detailed descriptions): Dryas heath Cassiope heath Graminoid heath Racomitrium moss Late snow-melt vegetation Dry moss tundra Moist moss tundra Graminoid wetland Brown moss wetland Biological crust Bare ground Rocks & gravel Water Snow The land cover map was created using a Random Forest (RF) classifier trained on in situ vegetation data and features derived from two UAS datasets: a digital surface model (DSM) and a multispectral orthomosaic (OM). The dataset used to train the RF classifier consisted of 156 in situ vegetation plots and 203 points added based on visual inspection of orthomosaics created with a DJI Mavic 3M. For each plot, we sampled all pixels with a pixel center within a 20 cm circular buffer around the plot center, which provided 4462 labeled observations. The predicted land cover classification map was post-processed using a 30 cm modal filter to remove noise. The land cover classification map (width: 13446, height: 18238) consists of one band: Land cover type: ranges from 1 to 14 (see README for land cover type descriptions) Digital surface model The DSM was created using data collected with a Qube 240 LiDAR sensor (Quantum-Systems GmbH, Gilching, Germany) acquired on 16.07.2024 at a flight altitude of 100 m (terrain following, average leg altitude) with 50% side overlap. The resulting LiDAR point cloud had a point density of about 94 points/m$^2$. The LiDAR point cloud was processed using post-processing kinematic (PPK) GNSS data from the ETPOS CORS network (Kartverket, Norway). We used Emlid Studio (version 1.9) to convert the GNSS observations to RINEX and subsequently created the smoothed best estimate of trajectory in Applanix POSPac UAV (version 9.0). Using YellowScan CloudStation (version 2403.0.1), we corrected the trajectory of the UAS, performed georeferencing, and created the DSM. Using the DSM we derived the following hydro-topographic features: Height above river Slope Sine and cosine of aspect Mean TPI within 3.1 m and 101.0 m windows Mean TWI within a 5.0 m window Multispectral orthomosaic The OM was created using images taken with a MicaSense AltumPT multispectral sensor (EagleNXT, Allen, TX, USA) and is a composite of three flights (13.-17.07.2024) at a flight altitude of 100 m (terrain following, average leg altitude) with 78% forward and 80% side overlap. The resulting ground sampling distance was about 2.1 cm. The multispectral images were processed in Agisoft Metashape (version 1.8.4) using a Structure-from-Motion workflow. The data was pre-processed by performing reflectance calibration with images of a calibrated reflectance panel (MicaSense CRP 2) and data from a downwelling light sensor (MicaSense DLS 2). Alignment between the OM and DSM was ensured through the use of manual reference points using additional orthomosaics created from multispectral data collected with a DJI Mavic 3M and processed using a PPK workflow. The original resolution of the OM was 5 cm per pixel, which we resampled to a resolution of 10 cm per pixel using mean aggregation. Using the OM we derived the following multispectral features: Red ($668 nm \pm 7 nm$), green ($560 nm \pm 13.5 nm$), blue ($475 nm \pm 16 nm$), red edge ($717 nm \pm 6 nm$), and near-infrared ($842 \pm 28.5 nm$) bands Standard deviation of red and near-infrared bands within a 1.1 m window Mean of Normalized Difference Red-Edge Index (NDRE) within a 0.5 m window Normalized Difference Vegetation Index (NDVI) Normalized Difference Water Index (NDWI) Standard deviation of NDVI and NDWI within a 3.1 m window Vegetation We collected in situ plant community composition data from 158 plots (50 cm $\times$ 50 cm) during the two field campaigns (2024: 52 plots, 2025: 106 plots). For each species present in a plot, we estimated the percentage coverage and measured the average vegetation height. We also performed soil moisture measurements (2024: 49 plots, 2025: 0 plots) with a handheld Delta-T Devices SM150T Soil Moisture Sensor (Delta-T Devices Ltd, Cambridge, United Kingdom) by measuring the center and corners of each plot. Additionally, we took photos of all plots and measured the plot location with an Emlid Reach RS2+/RS GNSS receiver (Emlid Tech Kft., Budapest, Hungary) with real-time kinematics (RTK) connection to the ETPOS CORS network (Kartverket, Norway). Accuracy assessment The land cover classification was evaluated using class-wise stratified 10-fold cross-validation, where the observations were grouped by plot in order to avoid spatial autocorrelation. To evaluate the model performance, we calculated the F1-score, which ranges from 0 to 1 and is defined as the harmonic mean of recall and precision. The RF classifier achieved a high overall performance with a class-weighted F1-score of 0.78, with the following performance per class: Land cover type Precision Recall F1-score Coverage of the study area Dryas heath 0.68 0.67 0.67 5.7 % Cassiope heath 0.68 0.78 0.73 6.5 % Graminoid heath 0.9 0.77 0.83 0.3 % Racomitrium moss 0.96 0.89 0.93 0.8 % Late snow-melt vegetation 0.77 0.69 0.73 9.9 % Dry moss tundra * 0.46 0.53 0.49 4.1 % Moist moss tundra * 0.48 0.63 0.55 1.2 % Graminoid wetland * 0.77 0.69 0.73 6.1 % Brown moss wetland 0.82 0.88 0.85 2.5 % Biological crust 0.86 0.82 0.84 14.9 % Bare ground 0.79 0.8 0.8 6.9 % Rocks & gravel 0.93 0.88 0.9 39.2 % Water 0.91 0.99 0.95 1.7 % Snow 0.93 0.95 0.94 0.1 % *Dry and Moist moss tundra have a lower F1-score in comparison to other land cover types due to misclassification as Graminoid wetland, so that these land cover types should be used with caution.