WCSE

Using Neural Networks for the Extraction of Built-Up Areas from Sentinel-2

Abstract— As urban areas develop, changes occur in their landscape. Cities or metropolitan areas are typically warmer with slightly higher temperatures compared to their adjacent rural areas. This temperature difference is due to the unusual state known as the urban heat island. Buildings, roads and other infrastructure replace open land and vegetation. Detecting urban areas from remote sensing images plays an important role in the field of Earth observation. In this paper, we propose to exploit the benefit of Sentinel-2 images to extract built-up areas. The following machine learning methods have been tested: random forests, support vector machines and multi layered perceptron. Experiments are performed in Ostrava area, in the Czech Republic. The validation is carried out using a control dataset which is Corine Land Cover 2012. The results were validated by a Kappa index and it showed that random forests is among the best performing method for the classification. The main result obtained in this paper is that the neural network considered here provides a satisfying effect for the classification of urban multispectral images.

Index Terms— Urban areas, Sentinel 2, Land cover, Machine learning, Neural networks

Lucie Orlíková
VSB – Technical University of Ostrava, CZECH REPUBLIC

[Download]

Cite: Lucie Orlíková, "Using Neural Networks for the Extraction of Built-Up Areas from Sentinel-2," Proceedings of 2018 the 8th International Workshop on Computer Science and Engineering, pp. 308-312, Bangkok, 28-30 June, 2018.