DamageTriage-Bench

Per-building post-disaster damage assessment benchmark for satellite imagery, with a fine-grained 5-class typology taxonomy (no damage, partial/total roof damage, partial/total structural damage) rather than the FEMA-style 4-class severity scale used by xBD.

Quick stats

Tiles	7,472 (1024 × 1024 PNG, sub-meter GSD)
Total building instances	≈ 76,500
Damage classes	5 (typology)
Disasters	3 (Hurricane Michael 2018, Hurricane Helene 2024, 2025 LA Palisades/Eaton wildfire complex)
Sub-events	12
Train / Val / Test split	5,229 / 1,120 / 1,123 tiles (stratified per sub-event)

Directory layout

.
├── images/                       # 7,472 post-event RGB tiles (1024×1024 PNG)
├── damage/                       # 7,472 per-tile unified RGB polygon masks
├── stratified_splits.json        # Train / Val / Test split (canonical)
├── class_index.json              # Per-class tile listings
└── README.md (this file)

images/<tile_id>.png and damage/<tile_id>.png are paired by filename.

Annotation format

Annotations are unified RGB polygon masks: a single 1024 × 1024 RGB mask per tile, where each pixel's colour identifies the damage class of the building instance it belongs to (or background).

Colour (R, G, B)	Class
(0, 0, 0)	Background / ignore
(255, 255, 255)	Undamaged
(0, 255, 83)	Partial Roof Damage
(246, 255, 11)	Total Roof Damage
(255, 138, 18)	Partial Structural Damage
(255, 0, 0)	Total Structural Collapse

Connected components of the same colour correspond to building instances. Polygons do not overlap.

Per-class instance counts (val / test)

Class	Val (n)	Val (%)	Test (n)	Test (%)
0. Undamaged	9,288	73.1	9,346	74.6
1. Partial Roof Damage	1,657	13.0	1,561	12.5
2. Total Roof Damage	165	1.3	145	1.2
3. Partial Structural Damage	543	4.3	471	3.8
4. Total Structural Collapse	1,062	8.3	999	8.0
Total	12,715	100	12,522	100

The training split follows the same long-tail distribution with the remaining ~60,000 instances.

Disasters and sub-events

12 acquisition sub-events span three disasters:

Sub-event	Hazard	Train	Val	Test	Total
wildfire_1	WF (LA)	220	47	47	314
wildfire_2	WF (LA)	208	45	44	297
wildfire_3	WF (LA)	92	20	20	132
wildfire_4	WF (LA)	191	41	41	273
wildfire_5	WF (LA)	201	43	43	287
wildfire_6	WF (LA)	86	18	19	123
wildfire_7	WF (LA)	146	31	32	209
wildfire_8	WF (LA)	99	21	22	142
hurricane_michael_2018	HUR	1,263	271	270	1,804
hurricane_helene_2024_v1	HUR	220	47	48	315
hurricane_helene_2024_v2	HUR	860	184	185	1,229
hurricane_helene_2024_late	HUR	1,643	352	352	2,347

The eight wildfire_* sub-events partition the 2025 Los Angeles Palisades / Eaton wildfire complex into spatially disjoint regions. The three hurricane_helene_2024_* sub-events correspond to separate Helene acquisitions. hurricane_michael_2018 covers the 2018-10-11 NOAA Emergency Response Imagery for Hurricane Michael.

Annotation rubric

Building polygons were manually annotated at the per-instance level by trained annotators following a damage-typology rubric jointly developed with structural-engineering domain experts.

Class 0 denotes buildings with no visible roof or structural damage. For damaged buildings, labels follow a two-step rule:

1. Does visible damage extend below the roof surface into structural components? If no, the instance is roof damage. If yes, it is structural damage.
2. Within each branch, a 50 % affected-area threshold separates partial from total. Area is estimated relative to the visible roof or building footprint.

Each polygon is assigned exactly one of the five typology classes.

Splits

stratified_splits.json defines a per-sub-event 70 / 15 / 15 split. Within each sub-event, tiles are partitioned so that every split has the same per-sub-event proportions as the full dataset. Splits are over tiles, not over building instances.

Test fold is never read during training or model selection.

import json
splits = json.load(open("stratified_splits.json"))
splits["seed"]                    # 42
splits["ratios"]                  # [0.7, 0.15, 0.15]
splits["all"]["train"][:5]        # first 5 train tile IDs
splits["events"][k]["test"][:5]   # first 5 test tile IDs from sub-event k

How to load

from datasets import load_dataset

# (After upload) the canonical loader.
ds = load_dataset("<your-hf-org>/DamageTriage-Bench")

For PyTorch training pipelines, the companion code repository /dinov3-damage-assessment provides a get_dataloaders() entry point that handles the unified-mask decoding and the stratified split.

Reproducibility

The full training recipe that produced the headline macro-F1 = 0.619 on the test split is documented in the companion code repository (AGENTS.md → §"v11 reference recipe").

License

CC BY-NC 4.0 — non-commercial research use only.

Acknowledgements

Imagery for the Michael event is sourced from the NOAA Emergency Response Imagery program. Imagery for the Helene and LA wildfire events is sourced from publicly released post-event capture flights; sources and ground-sampling distance are listed per-event in the companion paper.

Citation

@article{damagetriage2026,
  title   = {Damage-TriageFormer: Post-Event Foundation Models for
             Decision-Relevant Building Damage Typology},
  author  = {Xiao, Yiming and Mostafavi, Ali},
  journal = {tba},
  year    = {2026},
}