Sub-pixel building area mapping based on synthetic training data and regression-based unmixing using Sentinel-1 and -2 data

The identification of buildings has become a major research focus of settlement mapping with Earth Observation data. Building area or building footprint data is particularly required in research related to population, such as disaster risk management or urban health. This study examined the suitability of machine learning regression-based unmixing for quantifying the pixel-wise share of building area with decametre resolution Copernicus Sentinel-1 and Sentinel-2 imagery. Compared to using a single-step approach directly estimating building area, leading to an over-estimation of building area compared to non-building impervious surface area due to feature similarity, the introduction of a hierarchical approach considerably improved mapping results. While the original mapping resolution was 10 m, we found that building area was most accurately mapped starting at a spatial resolution of 100 m – a resolution well suitable for many urban applications. The proposed approach is widely transferable in space as it used spatially robust spectral-temporal metrics from time series imagery and as its requirements for training data are very limited.