U-Net feature fusion for multi-class semantic segmentation of urban fabrics from Sentinel-2 imagery: an application on Grand Est Region, France

ABSTRACT

Urban areas are increasing since several years as a result of development of built-up areas, network infrastructure, industrial areas or other built-up areas. This urban sprawl has a considerable impact on natural areas by changing the functioning of ecosystems. Mapping and monitoring Urban Fabrics (UF) is therefore relevant for urban planning and management, risk analysis, human health or biodiversity. For this research, Sentinel-2 (level 2A) single-date images of the East of France, with a high spatial resolution (10 m), are used to assess two semantic segmentation networks (U-Net) that we combined using feature fusion between a from scratch network and a pre-trained network on ImageNet. Moreover three spectral or textural indices have been added to the both networks in order to improve the classification results. The results showed a performance gain for the fusion methods in classifying several UF. However, there is a difference in performance depending on the urbanization gradient; highly urbanized areas provide a better distinction of some UF’s classes than rural areas.

Acknowledgements

All the network figures were designed using PlotNeuralNet (Iqbal 2018). We thanks to the Spatial Data Infrastructure GeoGrandEst provided the reference data used in this study and the Theia Services and Data Infrastructure for the Sentinel-2A imagery.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Notes

1. http://www2.isprs.org/commissions/comm3/wg4/semantic-labeling.html.

2. https://hyperspectral.ee.uh.edu/?page_id=459.

3. https://www.datagrandest.fr/portail/fr/projets/occupation-du-sol.

4. https://geoservices.ign.fr/documentation/donnees/vecteur/bdtopo.

5. https://numpy.org/.

Additional information

Funding

This research is part of the PhD Thesis supported by the French funded project ANR TIMES ‘High-performance processing techniques for mapping and monitoring environmental changes from massive, heterogeneous and high frequency data times series’ [ANR-17-CE23-0015] and by the French TOSCA project AIMCEE [CNES, 2019-2022].