Work

De la segmentació semàntica a petjades d'edificis per a SIG: extracció i vectorització d'edificis a partir d'imatges aèries mitjançant aprenentatge profund

Year: 2026
Type: dissertation
Abstract: [EN] This Master s thesis proposal focuses on the automatic extraction of building footprints from high-resolution aerial imagery, with the goal of producing geospatial outputs that can be directly used in Geographic Information Systems. The proposed work addresses the full transition from pixel-level semantic segmentation to GIS-ready vector building footprints. To achieve this, the project will design and evaluate an end-to-end pipeline that starts from aerial images and reference building masks, trains deep learning models for building segmentation, applies post-processing to improve the quality of the predicted masks, and converts the final raster predictions into georeferenced polygon geometries. The experimental component will focus on encoder-decoder architectures commonly used in semantic segmentation, including U-Net and related convolutional models with pretrained encoders. The models will be trained on tiled aerial imagery and evaluated using standard segmentation metrics such as Intersection over Union, Dice/F1 score, precision, recall, accuracy, and boundary-based measures. Data augmentation will be used to improve robustness to variations in orientation, scale, illumination, roof appearance, and urban morphology. Beyond obtaining accurate binary masks, the project emphasizes the generation of practical GIS deliverables. Predicted probability maps will be converted into binary masks through thresholding and refined using image post- processing operations such as small-component removal and morphological filtering. Building contours will then be extracted, simplified, and transformed into polygon geometries. These geometries will be exported in geospatial formats such as GeoJSON, preserving the spatial reference information required for visualization, editing, and integration into GIS workflows. The expected contribution of the work is a complete and reproducible workflow that connects deep learning-based aerial image analysis with geospatial vectorization.
Language: English
Open Access status: green
OpenAlex ID: W7219273378
Cites:
Cited by:
Related to: 10