Towards Robust Road Segmentation in Aerial and Satellite Imagery with High- to Low-fidelity Labels

Kurzfassung

Roads are a critical part of our countries' infrastructure, one which evolves constantly under the pressure of urban expansion, enables rapid transportation of individuals and goods throughout the world every day, and is therefore in need of perpetual maintenance. However, such a complex network cannot be fully comprehended or even used efficiently without continuously mapping every road worldwide as they grow, a task that can hardly be accomplished by local authorities via ground surveys. In this respect, earth observation offers an essential opportunity by delivering a bird's eye view of entire cities or regions from around the globe thanks to satellite and aerial imagery. This approach made it possible to accelerate mapping efforts through crowd-sourced projects like OpenStreetMap, it still relies on timely and accurate contributions from volunteer annotators. Coincidentally, deep learning has made it possible to create models capable of learning from vast amounts of remote sensing data and the corresponding road annotations. In this thesis, I introduce an array of neural network-based methods to support diverse downstream tasks like urban monitoring and humanitarian relief missions with automatically-generated road maps. To let models learn to segment roads from existing large-scale road extraction datasets, it is necessary to make them robust towards ubiquitous annotation inaccuracies, so-called label noise, as human annotators are not infallible. I first focus on training convolutional neural networks to be resilient to such errors by using all the valid labels while learning to ignore the wrong ones. While intentionally feeding the models with an increased level of noise and letting it explicitly distrust the labels, I find out that they perform better under such adversarial conditions than if they were left unaware of the low level of label noise present in the original labels. Approaching the problem from the opposite angle, I study the cost of producing high-fidelity labels for a limited amount of images, and their impact on the training, showing that few well-annotated images suffice to reach a high level of accuracy. Finally, leveraging all available high-fidelity annotations from public datasets and generating pseudo-labels for new regions thanks to the Segment Anything Model, I demonstrate the capability of vision transformers to learn to extract roads from vastly different sources of imagery and generalize with no additional labeling efforts to many areas around the world.