Comparing map-based and library-based training approaches for urban land-cover fraction mapping from Sentinel-2 imagery

Abstract

An improved trade-off between resolution, coverage and revisit time, makes Sentinel-2 multispectral imagery an interesting data source for mapping the composition and spatial-temporal dynamics of urban land cover. To fully realize the potential of Sentinel-2′s high amount of available data, efficient urban mapping workflows are required. Machine learning regression is a powerful approach to produce subpixel land cover fractions from remote sensing imagery, yet it requires mixed spectra for model training for which the fractions of the land cover classes present in the pixel are known. Typically, this data is obtained by sampling spectra from the image to be unmixed, and corresponding land-cover fractions from higher-resolution land cover reference data, i.e. map-based training. We propose synthetic mixing of library spectra as an alternative for producing land cover fraction training data for regression modelling, i.e. library-based training. The approach is applied to a Sentinel-2 image of the city of Brussels (Belgium) and part of its urban fringe for mapping Vegetation, Impervious, and Soil (VIS) fractions at 20 m resolution. VIS fraction maps obtained with library-based training have mean absolute errors below 0.1 for all three surface types. The composition of these three key surface categories and their spatial distribution is well represented for the entire area in resulting maps. As a proof of concept, library-based training is compared with the map-based training approach. The more flexible library-based training not only achieves similar mapping accuracies, but in most cases, outperforms the map-based training approach in terms of bias and magnitude of error. The outcome of the research suggests that use of spectral libraries and synthetic mixing may provide an efficient modelling framework for regression-based mapping from Sentinel-2 imagery in operational contexts.