| Keywords [eng] |
Semantic segmentation, Automated annotation, Satellite imagery, Segment Anything Model, Segment Anything Model in High Quality, Lithuanian topography, Remote sensing, Deep learning |
| Abstract [eng] |
This master’s thesis investigates methodologies for the semantic segmentation of satellite images to achieve automated annotation, focusing on Lithuania’s different landscapes - Kėdainiai, Varėna, Klaipėda and Kaunas regions. Research evaluates the efficacy of segmentation models, specifically the Segment Anything Model (SAM) and its high-quality variant (SAM-HQ), with various Vision Transformer (ViT) checkpoints (ViT-B, ViT-L, ViT-H). Data retrieved from Sentinel-2 satellite imagery are processed for use, including cloud removal using the UnCRtainTS tool, and analyzed using the SamGeo Python package. The study compares model performance through both input prompt segmentation and automatic mask generation, with model’s predicted Intersection over Union (IoU) scores as the primary metric for the latter. Findings indicate that SAM-HQ models generally outperform SAM with the ViT-H checkpoint often yielding the highest accuracy. However, performance varies with landscape characteristics - agriculturally diverse Kėdainiai region showed higher segmentation confidence compared to the densely forested Varėna region. Input prompt segmentation proved less reliable for consistent automatic annotation. The research concludes that while SAM-HQ offers a robust way for automated annotation, model and parameter selection are crucial and landscape dependent. The work presents workflow steps and region specific insights for applying segmentation models on Lithuanian geospatial data. |
