kshitijrajsharma/dinov3-damage-assessment
Building-level disaster damage assessment. A frozen DINOv3 ViT-L/16 satellite backbone with a
trainable UperNet decoder and two heads (localization + ordinal 4-class damage). Given building
footprints and post-disaster imagery (optionally pre-disaster imagery), it assigns each building a
damage level on the xBD Joint Damage Scale with a calibrated confidence.
Damage classes
no-damage, minor-damage, major-damage, destroyed.
Object-level metrics (val split, 6166 buildings)
| class | F1 |
|---|
| no-damage | 0.923 |
| minor-damage | 0.5797 |
| major-damage | 0.6623 |
| destroyed | 0.8885 |
Macro F1 0.7634 | harmonic damage F1 0.7348.
The damage F1 is computed on building pixels only (background excluded). Numbers are
in-distribution (the xView2 benchmark splits by tile, so the same events appear in train and
test); cross-event generalisation to a fully unseen disaster is harder for the subtle
minor/major classes.
Inputs and outputs
- Input: building footprints (GeoJSON) + post-disaster RGB GeoTIFF, optionally a pre-disaster
GeoTIFF aligned to the post grid. Footprints must overlay the post image correctly.
- Output: the footprints annotated with
damage_class, damage, confidence, review, and
per-class probabilities.
Files
model.onnx: self-contained inference graph (post, pre -> damage logits).model.ckpt: Lightning checkpoint for evaluation or further training.config.yaml: training configuration.calibration.json: confidence temperature (1.1719).
Confidence
Apply softmax(logits / 1.1719) for calibrated probabilities. Buildings below the
confidence threshold are flagged for human review.
Backbone
DINOv3 ViT-L/16 (sat493m), frozen. Decoder, fusion, and heads are the only trained parameters
(~24M). Trained on xBD (xView2), license CC BY-NC-SA 4.0; this model inherits the non-commercial
share-alike terms.
