ChrisRPL/satellite-disruption-triage-v0

Satellite Disruption Triage Dataset (v0)

TL;DR

This is a small, high-quality auxiliary dataset of paired satellite images (baseline + current) from 11 global disaster events, annotated with structured JSON triage outputs for training or evaluating vision-language models (VLMs) on macro-scale civilian disruption detection.

• 120 examples total: 92 train, 28 eval
• Source: BRIGHT dataset (XView2-compatible format derived from Kullervo/BRIGHT)
• Purpose: Structured VLM fine-tuning and evaluation for satellite-based humanitarian disruption triage
• License: CC-BY-NC-4.0 (Maxar Open Data via BRIGHT)

---

What this dataset is

Each example contains:

1. A baseline satellite image of a location before a disaster event
2. A current satellite image of the same location after the event
3. A structured JSON output with exactly these fields:

Field	Type	Description
action	string	One of: discard, defer, downlink_now
category	string	Short disruption type label (e.g., earthquake_building_damage)
rationale	string	One or two sentence operational explanation
bbox_norm	float[4] or null	Normalized bounding box [x_min, y_min, x_max, y_max] of affected area, or null if no disruption

Action definitions

• discard — No meaningful macro-scale civilian disruption visible. Buildings and infrastructure appear structurally intact in the current image compared to baseline.
• defer — Minor to moderate civilian infrastructure disruption visible. Damage is present but localized; may warrant later review but does not indicate immediate widespread impact requiring urgent action.
• downlink_now — Significant macro-scale civilian disruption visible (widespread building damage/destruction, flooding, wildfire destruction, etc.). Urgently warrants satellite downlink and humanitarian response coordination.

Category taxonomy

Category	Description	Example sources
no_disruption	No visible damage (paired with discard)	Any event
earthquake_building_damage	Seismic damage to civilian buildings	Turkey, Morocco, Noto, Haiti
wildfire_structure_damage	Wildfire destruction of civilian structures	Hawaii, Marshall Fire
volcanic_eruption_damage	Lava flow or ashfall damage to civilian structures	La Palma, Congo
flood_infrastructure_damage	Flood-related damage to buildings and infrastructure	Libya
civilian_explosion_damage	Civilian building damage from industrial/urban explosions	Beirut, Bata

Excluded categories: No explicit hospital/humanitarian site labels, no road-access labels, no military tactical data. The dataset is scoped to civilian macro-scale building disruption only.

---

What this dataset is NOT

1. Not a military targeting dataset — No military objectives, strike damage, or tactical analysis labels.
2. Not a real-time monitoring claim — Labels are based on post-event archived satellite imagery and do not imply real-time detection capability.
3. Not a tiny-object detection dataset — Focus is on macro-scale (tile-level or large-cluster) disruption, not individual small-object detection.
4. Not a hospital/humanitarian-site-specific dataset — While we prefer civilian-relevant labels, source data lacks explicit hospital or humanitarian site annotations.
5. Not exhaustively complete — This is version 0 (~120 examples) intended as an auxiliary fine-tuning/evaluation resource, not a production-scale production dataset.

---

Source datasets used

Primary source: BRIGHT (XView2 format)

• HF repo: GabeT29/BRIGHT-XView2Format
• Parent repo: Kullervo/BRIGHT
• Paper: Chen et al., "BRIGHT: a globally distributed multimodal building damage assessment dataset with very-high-resolution for all-weather disaster response", Earth System Science Data, 17, 6217–6253, 2025. DOI: 10.5194/essd-17-6217-2025
• License: CC-BY-NC-4.0 (Maxar Open Data)

BRIGHT provides paired pre-event optical and post-event SAR images with pixel-level damage masks:

• 0 = background (no building)
• 1 = intact
• 2 = damaged
• 3 = destroyed

We mapped these mask-derived damage statistics to our triage schema.

Datasets inspected but not used as primary source

Dataset	Why excluded / limited use
xBD / xView2	Official download requires registration; HF mirrors incomplete. BRIGHT supersedes for our use case.
SpaceNet 8	Focused on flood detection + road network; requires tortilla library for HF version. Road labels did not cleanly map to our macro-scale building triage schema.
KOlegaBB/damage_assessment_ukraine	Small (2,219 instances), uses Google Maps imagery with redistribution restrictions. BRIGHT includes Ukraine/Myanmar conflict tiles under Maxar Open Data license, making BRIGHT preferable.
Sen1Floods11	No pre-event baseline imagery; SAR-only without paired optical. Labels are pixel-level flood masks without building-level damage severity. Does not cleanly support our paired-image macro-scale triage task.

---

Event provenance

Event	Examples	Region	Year	Type
Turkey earthquake	22	Turkey	2023	Natural disaster
Marshall wildfire	20	USA	2022	Natural disaster
Noto earthquake	17	Japan	2024	Natural disaster
La Palma volcano	14	Spain	2021	Natural disaster
Hawaii wildfire	9	USA	2023	Natural disaster
Libya flood	8	Libya	2023	Natural disaster
Beirut explosion	7	Lebanon	2020	Man-made disaster
Congo volcano	7	DRC	2021	Natural disaster
Bata explosion	7	Equatorial Guinea	2021	Man-made disaster
Morocco earthquake	6	Morocco	2023	Natural disaster
Haiti earthquake	3	Haiti	2021	Natural disaster

Coverage: 11 events across 4 continents, 5 disaster types (earthquake, wildfire, volcanic eruption, flood, explosion), natural and man-made origins.

---

Labeling methodology

Labels were derived algorithmically from BRIGHT pixel-level damage masks with the following pipeline:

1. Damage statistics: For each 1024×1024 tile, we computed the ratio of damaged+destroyed pixels to total building pixels from the mask.
2. Triage mapping:
   • damage_ratio < 0.05 → discard
   • 0.05 ≤ damage_ratio < 0.35 AND destruction_ratio < 0.15 → defer
   • Otherwise → downlink_now
3. Category: Inferred from BRIGHT event name (e.g., "turkey-earthquake" → earthquake_building_damage).
4. Rationale: Template-generated based on action + category, written in operational language.
5. Bounding box: Computed as the axis-aligned bounding box of all damaged/destroyed pixels, normalized to [0, 1] per tile. null for discard examples.
6. Deduplication: Perceptual hash (16×16 image thumbnail) on post-event tiles to exclude near-identical examples.
7. Diversity enforcement: Round-robin selection across events to prevent single-event dominance.

---

Splits

Split	Count	Description
train	92	Stratified by action (27 discard, 27 defer, 38 downlink_now)
eval	28	Stratified by action (8 discard, 8 defer, 12 downlink_now)

Stratification ensures both splits contain representative proportions of all three triage actions.

---

File formats

This repository provides two machine-readable formats describing the same examples:

1. Flat JSONL (train_flat.jsonl, eval_flat.jsonl)

For structured evaluation and non-conversational training:

{
  "example_id": "bright_turkey-earthquake_val_images_turkey-earthquake_00000299",
  "baseline_image": "images/baseline/bright_turkey-earthquake_val_images_turkey-earthquake_00000299_baseline.png",
  "current_image": "images/current/bright_turkey-earthquake_val_images_turkey-earthquake_00000299_current.png",
  "target_output": {
    "action": "downlink_now",
    "category": "earthquake_building_damage",
    "rationale": "Widespread building collapse visible in civilian area following earthquake. Multiple structures destroyed, indicating potential mass casualty event requiring urgent response.",
    "bbox_norm": [0.3203, 0.2676, 0.8291, 0.7568]
  },
  "source_dataset": "BRIGHT (GabeT29/BRIGHT-XView2Format)",
  "source_event": "turkey-earthquake",
  "source_image_name": "val/images/turkey-earthquake_00000299",
  "provenance": "BRIGHT dataset, event=turkey-earthquake, tile=val/images/turkey-earthquake_00000299, source_split=val",
  "damage_ratio": 0.4521,
  "destruction_ratio": 0.2014
}

2. SFT-ready messages format (train_sft.jsonl, eval_sft.jsonl)

For vision-language model fine-tuning in standard conversation format:

{
  "example_id": "bright_turkey-earthquake_val_images_turkey-earthquake_00000299",
  "images": [
    "images/baseline/..._baseline.png",
    "images/current/..._current.png"
  ],
  "messages": [
    {
      "role": "system",
      "content": "You are a satellite imagery analyst performing structured disruption triage. You are given two satellite images of the same location: a baseline image from an earlier date and a current image from a later date. Compare the two images and determine whether there is meaningful macro-scale civilian disruption visible in the current image. Return your assessment as a JSON object with exactly these fields: \"action\" (one of \"discard\", \"defer\", or \"downlink_now\"), \"category\" (short disruption type label), \"rationale\" (one or two sentence operational explanation), \"bbox_norm\" (null or [x_min, y_min, x_max, y_max] normalized bounding box of affected area)."
    },
    {
      "role": "user",
      "content": "Compare these two satellite images of the same location.\n\nImage 1 (Baseline): <image_baseline>\nImage 2 (Current): <image_current>\n\nAnalyze whether there is meaningful macro-scale civilian disruption visible in the current image compared to the baseline. Return your assessment as a strict JSON object."
    },
    {
      "role": "assistant",
      "content": "{\"action\":\"downlink_now\",\"category\":\"earthquake_building_damage\",...}"
    }
  ],
  "source_dataset": "BRIGHT (GabeT29/BRIGHT-XView2Format)",
  "provenance": "BRIGHT dataset, event=turkey-earthquake, tile=..."
}

Note: Image paths in messages are referenced via the images list. The placeholder tokens <image_baseline> and <image_current> in user content should be replaced with actual image tokens by the training pipeline.

---

Intended use

• VLM SFT fine-tuning — Train small-to-medium VLMs to output structured JSON from paired satellite images.
• Structured evaluation — Benchmark a VLM's ability to correctly classify triage actions and localize affected areas.
• Prompt engineering research — Study how system prompts and image ordering affect structured output quality for geospatial analysis.
• Auxiliary dataset — Combine with larger general-purpose instruction datasets or domain-specific satellite datasets.

---

Main limitations

1. Small size (120 examples) — This is intentionally a high-quality but small v0 dataset. It is suitable for few-shot fine-tuning, evaluation, or as an auxiliary component in a larger training mix, but not sufficient as a standalone training corpus for a full model.

2. Algorithmic labels with template rationales — Triage actions are derived from pixel-level mask statistics using fixed thresholds, not human expert judgment. Rationale text is template-generated rather than written by domain experts. A production system should include human-in-the-loop validation or at least expert spot-checking.

3. No explicit hospital / humanitarian site / road-access labels — The source BRIGHT dataset provides building-level damage masks but does not label hospitals, humanitarian sites, or road passability. Categories like "hospital disruption" or "major road-access disruption" referenced in the task specification are not present in this v0 dataset. Adding these would require overlaying OpenStreetMap amenity data or using a complementary dataset like SpaceNet 8 in a future version.

4. License restrictions — Derived from Maxar Open Data under CC-BY-NC-4.0. Not for commercial use without additional licensing from Maxar.

5. Image modality gap — BRIGHT pairs pre-event optical (RGB) with post-event SAR (radar). The baseline and current images are different sensor modalities, which may create artifacts unrelated to actual change. This is documented and expected, but VLM training pipelines should account for modality-aware prompting or multimodal understanding.

---

How to use

from datasets import load_dataset

# Load the flat evaluation format
ds = load_dataset("ChrisRPL/satellite-disruption-triage-v0", data_files="train_flat.jsonl", split="train")

# Or load the SFT format
ds_sft = load_dataset("ChrisRPL/satellite-disruption-triage-v0", data_files="train_sft.jsonl", split="train")

# Load paired images
from PIL import Image
import json

example = ds[0]
baseline = Image.open(example["baseline_image"])
current = Image.open(example["current_image"])
target = json.loads(example["target_output"])

---

Citation

If you use this dataset, please cite both this dataset and the original BRIGHT paper:

@dataset{satellite_disruption_triage_v0,
  title = {Satellite Disruption Triage Dataset v0},
  author = {Hugging Face Agent},
  year = {2025},
  url = {https://huggingface.co/datasets/ChrisRPL/satellite-disruption-triage-v0}
}

@article{Chen2025Bright,
  AUTHOR = {Chen, H. and Song, J. and Dietrich, O. and Broni-Bediako, C. and Xuan, W. and Wang, J. and Shao, X. and Wei, Y. and Xia, J. and Lan, C. and Schindler, K. and Yokoya, N.},
  TITLE = {BRIGHT: a globally distributed multimodal building damage assessment dataset with very-high-resolution for all-weather disaster response},
  JOURNAL = {Earth System Science Data},
  VOLUME = {17},
  YEAR = {2025},
  NUMBER = {11},
  PAGES = {6217--6253},
  DOI = {10.5194/essd-17-6217-2025}
}

---

Version history

• v0.1.0 (2025-04-23) — Initial release. 120 examples from 11 BRIGHT events. 92 train / 28 eval.