Upscaling vegetation cover from UAV to satellite imagery

Abstract. This study proposes an upscaling approach to estimate multi-species fractional cover of coastal dune vegetation by combining UAV-based (Uncrewed Aerial Vehicles) classification and PlanetScope imagery. A thirteen-species classification generated from a UAV multispectral dataset at 3cm spatial resolution was used to derive the Fractional cover (FC) as species composition on PlanetScope dataset. The spectral representativeness and consistency of the FC were tested, and then used as training information for a Dirichlet-based neural network, which provides both predictions and associated uncertainty. The method is tested on three coastal areas (200m along the coast) using a Leave-One-Area-Out (LOAO) validation scheme and then applied along the entire dune system of the Massacciucoli Regional Park (Tuscany, Italy). Results show stable performance across the three LOAO cycles and macro-RMSE values between 0.11 and 0.41. Herbaceous species show relatively low errors, while woody species and the sand background class remain more difficult to model, mainly due to structural complexity and spectral heterogeneity. Specific abundance-corrected metrics were computed to deal with rare species, whose evaluation is affected by near-zero values, showing robust results for most species.