The question.
A model that works on one aerial dataset may fail under a different resolution, season, geography, or sensor.
The research expands the earlier GeoAI concept into a comparative study of supervised segmentation, self-supervised representations, and foundation-model adaptation.
The proposed approach.
- Preserve georeferencing during tiling, sampling, and augmentation
- Compare task-specific models with adapted pretrained representations
- Use geography-separated evaluation to reduce spatial leakage
- Reproject predictions with uncertainty indicators for human GIS review
Target architecture.
Aerial imageryGeospatial tilingRepresentation adaptationSegmentation / change mapsUncertaintyGIS review
The difficult parts.
- Limited high-quality annotations
- Spatial leakage and sensor domain shift
- Ambiguous boundaries and changing ground truth
How to evaluate it.
- Segmentation and change-detection quality on held-out regions
- Performance under sensor, season, and resolution shifts
- Uncertainty calibration and annotation-effort trade-offs
Planned deliverables.
- A geography-aware benchmark specification
- A representation and adaptation comparison
- A GIS-ready output and uncertainty format
Design principles.
Hold out places, not just image tiles.
A map-ready prediction must preserve its spatial context.