tugcantopaloglu/solarscanner-solars

SolarScanner • U‑Net + ViT for Building Segmentation & Damage Classification

Model Details

Stage	Backbone	Dataset	Metric
Segmentation	U‑Net (ResNet‑50 encoder)	SpaceNet v2	IoU 0.766
Damage CLS	ViT‑B/16	xBD	Acc 0.856

Usage

from solars import load_seg_model, load_dmg_model
mask = load_seg_model().predict("image.tif")
labels = load_dmg_model().predict_patches("image.tif", mask)

Intended Use

Rapid mapping after earthquakes, floods, conflicts. Not for safety‑critical decisions without human review.

Limitations

City bias (4 training cities), damage‑class imbalance, RGB‑only.

Training

See GitHub repo for configs. AdamW, FP16, cosine schedule. (https://github.com/tugcantopaloglu/solarscanner-solars-paper-deep-learning)

Results

Task	Score
IoU	0.766
Acc	0.856

Citation

@unpublished{topaloglu2025solars,
  author  = {Tuğcan Topaloğlu},
  title   = {{SolarScanner}: Two‑Stage Deep Learning for Post‑Disaster Building Damage Assessment},
  year    = {2025}
}