Estimating Sub-pixel Vegetation Fractions in Vineyards Using a Multi-scale Hyperspectral UAV–PRISMA Framework

Hyperspectral satellite imagery provides valuable spectral information for vegetation monitoring, but its spatial resolution often limits detailed sub-pixel analysis in structured crops. This limitation is particularly relevant in vineyards, where satellite pixels may contain mixtures of vine canopy, inter-row soil, shadows, and other land-surface components. This study presents a preliminary multi-scale assessment of the extent to which vegetation fractions can be estimated from pan-sharpened PRISMA hyperspectral imagery using unmanned aerial vehicle (UAV) data as a high-resolution reference. Very-high-resolution UAV hyperspectral imagery acquired using a SPECIM AFX-10 sensor was used to derive reference fractions of vegetation and soil cover within a vineyard. These fractions were aggregated to the pan-sharpened PRISMA grid and used as reference data for estimating vegetation cover at the satellite scale. The analysis examined PCA-derived spectral components and vegetation indices to evaluate the spectral signal contained in mixed PRISMA pixels. Subsequently, several estimation approaches were compared, including Random Forest regression, Support Vector Regression, a shallow neural network, and a compact constrained neural model integrating PCA-reduced spectra, abundance constraints, and simple spatial descriptors. The results show that PRISMA VNIR bands retained a moderate cross-scale signal correlated with UAV-derived vegetation cover fractions. The compact constrained neural model achieved the best overall performance, with RMSE = 0.06 and R² = 0.70 for vegetation cover fraction estimation. Spatial error analysis revealed lower errors for mixed pixels located in the interior parts of the vineyard, while higher errors were observed in vegetation-dominant and peripheral pixels. The findings suggest that pan-sharpened PRISMA imagery can support sub-pixel vegetation cover estimation in structured crops while also highlighting uncertainties related to scale mismatch and the pan-sharpening process.