Abstract:
Roads found in informal settlements arise out of convenience are often not recorded or maintained by
authorities. This may cause issues with service delivery, sustainable development and crisis mitigation,
including COVID-19. Therefore, the aim of extracting informal roads from remote sensing images is of
importance. Existing techniques aimed at the extraction of formal roads are not completely suitable for the
problem due to the complex physical and spectral properties that informal roads pose. The only existing
approaches for informal roads, namely [62, 82], do not consider neural networks as a solution. Neural
networks show promise in overcoming these complexities due to the way they learn through training.
They require a large amount of data to learn, which is currently not available due to the expensive and
time-consuming nature of collecting such data sets. A problem that has been shown to come up when
working with computer vision data sets is data set bias. Data set bias adds to the already existing problem
of machine learning algorithms called overfitting. This paper implements a neural network developed for
formal roads to extract informal roads from three data sets digitised by this research group to investigate
the presence of data set bias. Three different geological areas from South Africa are digitised. We
implement the GAN-UNet model that obtained the highest F1-score in a 2020 review paper [1] on the
state-of-the-art deep learning models used to extract formal roads. We present quantitative and qualitative
results that concludes the presence of data set bias. We then present further work that can be done to
create a robust training data set for the development of an automatic informal road extraction model.
