Urban land cover mapping with Sentinel-2: a spectro-spatio-temporal analysis

Abstract

Because of its high spatial resolution (10 m and 20 m), rich spectral information (10 spectral bands), and short revisit period (5 days), Sentinel-2 provides new opportunities for earth observation. However, in urban environments, data limitations in the spatial and spectral dimensions constrain Sentinel-2’s performance: (i) the spatial resolution of Sentinel-2 is not sufficient to resolve the heterogeneous urban landscapes, and the prevalence of mixed pixels lowers the performance of image classifiers; (ii) the spectral resolution of Sentinel-2 is not able to fully account for the spectral variability of urban surface materials, which increases the probability that image classifiers mislabels urban land covers. Also, the 5-day temporal resolution makes Sentinel-2 image data suitable for time series analysis, but its contribution to urban land cover mapping still needs to be quantified. This study evaluated Sentinel-2’s performance in mapping urban land covers by mitigating the effect of spectral variability (using FDA, Fisher Discriminant Analysis), improving the spatial resolution of images (using UnFuSen2, a state-of-art Sentinel-2 image fusion approach), and utilizing temporal and spectral characteristics from image time series. Overall, we found that the image time series processed by UnFuSen2 enables the classifiers of k nearest neighbor (KNN), maximum likelihood (MLC), and random forests (RF) to perform the best, and multiple endmember spectra mixture analysis (MESMA) is suitable for classifying image time series that have been jointly processed by FDA and UnFuSen2. Besides, we found a significant contribution of spring and summer imagery to the improvement of land cover mapping accuracy in the Brussels Capital Region.