An Ensemble-Based Approach with Enhanced Loss Functions for Semantic Road Segmentation

Accurately analyzing road continuity in satellite imagery is a significant challenge, particularly when roads are obstructed by trees, shadows, or structures, complicating the identification of directional vectors of road sections. While most current research focuses on optimizing deep learning architectures, segmentation performance also heavily depends on the choice of loss function. Within the field of research, however, loss functions specifically designed for road segmentation tasks remain largely underexplored. In this thesis, we address these challenges by introducing an ensemble of convolutional neural network architectures for road segmentation, trained with different loss functions to enhance segmentation performance. Our approach leverages the strengths of various network designs and examines the impact of combining standard loss functions with novel variations inspired by topological analysis. By employing this methodology, we achieve state-of-the-art results across multiple benchmark datasets. Additionally, we provide a detailed investigation into the role of loss functions in road extraction, emphasizing their contribution to maintaining road continuity in challenging scenarios. This study offers valuable insights into designing loss functions tailored for road segmentation tasks and opens new avenues for future research in this field.