Comparing Global Sentinel-2 Land Cover Maps for Regional Species Distribution Modeling

Mapping the spatial and temporal dynamics of species distributions is necessary for biodiversity conservation land-use planning decisions. Recent advances in remote sensing machine learning have allowed high-resolution distribution modeling that can inform landscape-level decision-making. Here we compare performance three popular Sentinel-2 (10-m) land cover maps, including dynamic world (DW), European (ELC10), (WC), predicting wild bee richness over southern Norway. The proportion grassland habitat within 250 m (derived from maps), along with temperature distance to sandy soils, were used as predictors both Bayesian regularized neural network random forest models. Models using DW performed best (RMSE = 2.8 ± 0.03; average standard deviation across models), followed by ELC10 2.85 0.03) WC 2.87 0.02). All satellite-derived maps outperformed a manually mapped Norwegian dataset called AR5 3.02 When validating model predictions against citizen science data on solitary occurrences generalized linear models, found (AIC 2278 4), 2367 3), 2376 3). While differences RMSE observed between models small, they may be significant when such are prioritize patches landscape subsidies or management policies. Partial dependencies our showed increasing positively associated richness, thereby justifying schemes aim enhance semi-natural habitat. Our results confirm utility supporting suggest there scope monitor changes time given dense series provided products DW.