Label noise tolerance of deep semantic segmentation networks for extracting buildings in ultra-high-resolution aerial images of semi-built environments

Freely available building maps of rapidly changing built and semi-built environments may contain label noise. When temporal correspondence between images and labels does not hold, the labels may be subject to incorrectly observed building instances. For example, in most growing semi-built environments, such as the Kutupalong mega-camp in Bangladesh, labels corresponding to a past date may not be updated or might not have been properly labelled, resulting in label noise. Tagging/labelling can be done either manually (by humans) or automatically (by a machine/model). We manually label images for our stricter evaluation regime, but a trained model can automatically label images without human supervision. Our best performing model generates labels which improve F1-score by 17.2% and improve Intersection-over-Union score by 23.2%, when compared to the fidelity of commonly used noisy labels. Our stricter evaluation regime reveals interesting insights about the paradoxical behaviour of deep neural networks in conjunction to label noise.

Automated summary

Building maps may have label noise which can be addressed through manual or automatic labeling. The best performing model in the study improved label scores and reduced label noise.