Unmixing-based Sentinel-2 downscaling for urban land cover mapping

Abstract

With the launch of Sentinel-2 new opportunities for large scale urban mapping arise. However, the spectral information embedded in the Sentinel-2 20 m spatial resolution bands cannot yet be fully explored in heterogeneous urban landscapes. The 20 m image pixels are often composed of different land covers, resulting in a difficult to interpret mixed pixel spectrum. Here, we propose an unmixing-based image fusion algorithm (UnFuSen2) that self-adapts to the spectral variability of varying land covers and improves the image fusion accuracy by constraining the unmixing equations on the basis of spectral mixing models and the correlation between spectral bands of coarse and fine spatial resolution, respectively. When compared to alternative state-of-the-art downscaling methods UnFuSen2 consistently showed the highest accuracy when applied across test sites in three different European cities (RMSEUnFuSen2 = 203 vs RMSEalternatives = [252, 337]). In a next step, we applied Multiple Endmember Spectral Mixture Analysis (MESMA) on the downscaled Sentinel-2 image cube (i.e. ten 10 m bands) to generate subpixel urban land cover fractions. We compared our MESMA results against the traditional MESMA output as applied on the original Sentinel-2 image cube (i.e. four 10 m bands and six 20 m bands) and tested its robustness against reference data obtained over all three study sites. Results revealed an average decrease in RMSE of respectively 18% and 8% for impervious surface and vegetation fractions when our approach was compared to the traditional MESMA outcomes.