Deep Learning Solutions for Semantic Segmentation of Topographic Airborne LiDAR Point Clouds
[eng] Airbome LiDAR has become a key technology for large-scale mapping, providing dense three-dimensional point clouds that capture both natural landscapes and human-made infrastructure. Transforming these raw data into semantic information, however, remains a challenge due to irregular sampling pa...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2026 |
| País: | España |
| Institución: | Universidad de Oviedo (UNIOVI) |
| Repositorio: | Dipòsit Digital de la UB |
| OAI Identifier: | oai:diposit.ub.edu:2445/227498 |
| Acceso en línea: | https://hdl.handle.net/2445/227498 http://hdl.handle.net/10803/696825 |
| Access Level: | acceso abierto |
| Palabra clave: | Aprenentatge profund Visió per ordinador Imatges satel·litàries Deep learning (Machine learning) Computer vision Remote-sensing images |
| Sumario: | [eng] Airbome LiDAR has become a key technology for large-scale mapping, providing dense three-dimensional point clouds that capture both natural landscapes and human-made infrastructure. Transforming these raw data into semantic information, however, remains a challenge due to irregular sampling patterns, strong class imbalance, and distribution shifts across regions and acquisition conditions. This thesis addresses these challenges by developing deep learning solutions tailored to the semantic segmentation of topographic airborne LiDAR point clouds, with an emphasis on operational applicability in national mapping workflows. We first introduce object-centric methods for detecting and segmenting vertical objects in cluttered scenes, proposing a constrained sampling that enhances the characterization of vertical structures embedded in vegetation. We then investigate training and inference procedures to handle variability in density and scale, demonstrating how inductive biases and uncertainty-based inference substantially improve robustness without requiring architectural modifications. To reduce reliance on costly manual annotation, we adapt self-supervised learning to airborne LiDAR data, showing that Barlow Twins pre-training improves downstream segmentation, particularly for underrepresented classes. Finally, we explore domain adaptation and incremental learning, integrating LoRA into PointNet++ to achieve parameter-efficient fine-tuning. We show how LoRA facilitates the addition of new semantic categories with minimal overhead. Beyond methodological advances, this research contributes new resources to the community, including the release of the TerLiDAR dataset. In addition, several of the proposed methods have also been transferred to productive workflows at the Institut Cartografíe i Geologic de Catalunya, where they support tasks such as refining digital surface models, detecting missing transmission towers, recovering filtered power lines, and classifying wind turbines. Together, these contributions demonstrate how deep learning can be scaled and adapted for reliable, real-world airborne LiDAR semantic segmentation, bridging the gap between research innovation and productive workflows. |
|---|