Regional and pan-arctic climate and ecosystem models rely on land cover estimates from remote sensing to estimate and predict associated ecosystem processes within the carbon cycle, the water and energy balance as well as vegetation-related parameters like plant productivity or biomass. Systematically operating satellite systems provide a regular coverage of arctic land cover but do not have sufficient resolution to resolve small-scale surface heterogeneity. To what degree this aggregation of land cover introduces error and uncertainty into the model outcome remains a crucial research question.This work presents a unique dataset of a northern permafrost landscape combining detailed ground based measurements of land cover and evapotranspiration (ET) with multispectral high-resolution remote sensing data. Aerial images with a resolution of 0.3m per pixel show the fractionated polygonal tundra landscape on Samoylov Island (Lena Delta, Sibiria) being composed of 47% dry polygonal rims, 30% wet depressed polygonal centres and 23% small ponds. CHRIS Proba pixels with a resolution of 17m represent a mixture of those landscape units. Our study focuses on the following questions: 1. Decomposing mixed pixels: How well do subpixel estimates of CHRIS Proba correspond to the associated land cover units?2. Finding the right scale: How do different levels of aggregations of land cover units effect the regionalization of ET?During 40% of the observation period in summer 2008, net radiation values > 50 Wm² lead to distinct evapotranspiration rates of dry and wet surfaces with 1.8mm/d and 0.8mm/d, respectively. Land cover classes therefore have to be considered when deriving a regionalized ET value, i.e. ET rates of the different land cover units are weighted with their corresponding surface ratio. Here, we focus on ET but proposed methods can also be applied to other parameters, e.g. carbon fluxes.