Tree species map of Switzerland

Dominant tree species map of Switzerland We created a tree species map of Switzerland for the dominant tree species in the forested areas. The spatial resolution of the map is 10 m and the coordinate system is ETRS89-extended / LAEA Europe (EPSG 3035). The map comprises Sentinel-2 index time series from the year 2020, a digital elevation model and species reference data from the Swiss National Forest Inventory. The map is available as raster (.tif) or vector dataset (.gpkg) In total, the following 15 species were mapped: *Abies alba*, *Acer pseudoplatanus*, *Alnus glutinosa*, *Alnus incana*, *Betula pendula*, *Castanea sativa*, *Fagus sylvatica*, *Fraxinus excelsior*, *Picea abies*, *Pinus cembra*, *Pinus mugo arborea*, *Pinus sylvestris*, *Quercus petraea*, *Quercus robur*, *Sorbus aucuparia*. Access will be granted upon request. -br/--br/- Approach -br/--br/- Data - Swiss National Forest Inventory Data (stand species with - 60 % dominance in upper canopy; on at least more than 9 plots dominant) - Sentinel-2 time series (2020, Indices: CCI, CIRE, NDMI, EVI, NDVI) - Digital elevation model (DEM) (swissalti3d, 5 m) - Biogeographical regions (Federal Office for the Environment FOEN) - Forest mask 2017 (Approach: Waser et al., 2015) -br/--br/- Modeling approach We identified the most meaningful variables that led to separation of the respective groups by using random forest models with a forward feature selection (Meyer et al., 2018; Ververidis & Kotropoulos, 2005). In this approach, the final random forest model is solely built from the selected meaningful variables. By identifying meaningful variables, we can determine which variables might influence the grouping. Further, to avoid overfitting and overly optimistic results, we applied 10-fold spatial cross-validation and put all pixels from a plot in the same spatial fold. The modeling was realized using the CAST package in R (Meyer et al., 2022), based on the well-known caret package (Kuhn, 2022). We used the ranger package in R (Wright & Ziegler, 2017) to implement the random forest models, due to its short computation time. -br/--br/- Training data for modeling - 295 Sentinel-2, DEM & Biogeographical variables - 10525 tree species pixels -br/--br/- Selected variables for final model 1. EVI of 2020.05.16 2. NDMI of 2020.03.12 3. CIRE of 2020.04.16 4. NDMI of 2020.07.05 5. CCI of 2020.05.11 6. dem 7. CCI of 2020.08.14 8. NDMI of 2020.08.24 9. CCI of 2020.12.22 10. NDMI of 2020.04.21 11. NDMI of 2020.11.17