A carregar...
Publicação
PATHFinder: A Foundation Model for Road Mapping in Support of United Nations Humanitarian Affairs
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 31 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Accurate road segmentation in remote and challenging terrains is critical for humanitarian missions, yet existing methods often struggle with occlusions, faint roads, and low-visibility conditions. This research addresses two key challenges: (1) enhancing the accuracy and generalization of road segmentation models using transfer learning, and (2) improving road connectivity detection and edge segmentation through connectivity structures
and a custom loss function. The proposed PATHFinder model leverages transfer learning by pretraining on large-scale road datasets and fine-tuning on domain-specific data, achieving a 62.6% MIoU on remote and challenging terrains. The integration of connectivity structures and a custom loss function further improved precision by 5.55%, particularly in complex road intersections and low-visibility scenarios. The key contributions of this research include: (1) the development of the PATHFinder model, a framework for road segmentation in occluded, faint, and cloudy conditions; (2) improved model accuracy and generalization through transfer learning; and (3) enhanced road connectivity and edge segmentation, enabling robust performance in diverse environments. The model’s practical applicability is demonstrated through its potential to support humanitarian
missions, by providing accurate road mapping in areas where existing methods fail. Future work could explore the integration of dilation as a post processing technique, improved preprocessing of OpenStreetMap (OSM) labels, and the use of higher-resolution imagery to further enhance performance. This research underscores the effectiveness of transfer learning and architectural innovations in advancing road segmentation for real-world applications, particularly in humanitarian contexts. Code will be released at: https://github.com/Oraegbuayomide10/PathFinder
Descrição
Dissertation submitted in partial fulfilment of the requirements for the Degree of Master of Science in Geospatial Technologies
Palavras-chave
Road Mapping Foundation Model Humanitarian Missions Occlusions Satellite Imagery PATHFinder United Nations Global Service Center (UNGSC)