ObsCrisis-Bench

A multimodal benchmark for evaluating large vision-language models on extreme weather event analysis tasks.

Dataset Description

ObsCrisis-Bench contains 4,202 VQA samples across 127 extreme weather events in 8 disaster categories, covering 61 countries. Each sample combines satellite multispectral imagery (AMSU-A, HIRS, MHS sensors) with optional weather station data, and requires models to perform risk assessment, type classification, timing prediction, and impact analysis.

Disaster Categories

Category	Events	VQA Samples	Subtypes
cold-wave	5	140	cold-wave
earthquake	10	417	ground movement, tsunami
flood	17	523	coastal flood, flash flood, general flood, riverine flood
heat-wave	5	165	heat-wave
mass movement (wet)	6	189	landslide (wet), mudslide
storm	55	1,852	blizzard/winter storm, derecho, extra-tropical storm, general storm, hail, lightning/thunderstorm, sand/dust storm, severe weather, storm surge, tornado, tropical cyclone
volcanic activity	15	450	ash fall, general activity, lava flow, pyroclastic flow
wildfire	14	466	forest fire, general wildfire, land fire
Total	127	4,202

Task Structure

Each VQA sample belongs to one of three task categories:

Task Category	Samples	Percentage
Early Warning	1,524	36.3%
Impact Assessment	2,465	58.7%
Recovery Assessment	213	5.1%

Task Types

Early Warning (t1, t8, t12):

• Risk Detection (381 samples): Assess the risk of an impending disaster
• Type Classification (381 samples): Identify the specific type of potential disaster
• Arrival Time Prediction (381 samples): Predict when the disaster will arrive
• Duration Prediction (381 samples): Predict how long the disaster will last

Impact Assessment (t1, t8, t12, t15, tmax):

• CPI / Crisis Prevention Index (610 samples): Estimate the crisis prevention index
• Total Affected (485 samples): Predict total affected population
• Total Deaths (455 samples): Predict total deaths
• No. Affected (400 samples): Predict number of affected people
• No. Injured (260 samples): Predict number of injured people
• Magnitude (185 samples): Predict earthquake magnitude
• No. Homeless (70 samples): Predict number of homeless people

Recovery Assessment (t15, tmax-1):

• Recovery Time Prediction (213 samples): Predict recovery time after the event

Timestep Convention

• t1 = 14 days before event start
• t8 = 7 days before event start
• t12 = 3 days before event start
• t15 = event start day
• tmax = maximum impact time (event-dependent)
• tmax-1 = one timestep before maximum impact
• t_number = (image_date - event_start_date).days + 15

Dataset Structure

ObsCrisis-Bench/
├── cold-wave/                          # Satellite imagery
│   └── cold-wave1/<event_id>/<sensor>/<band>/t1.png ... tN.png
├── cold-wave_json/
│   └── All.json                        # VQA records
├── heat-wave/
├── heat-wave_json/All.json
├── earthquake/
├── earthquake_json/All.json
├── flood/
├── flood_json/All.json
├── mass movement (wet)/
├── mass movement (wet)_json/All.json
├── storm/
├── storm_json/All.json
├── volcanic activity/
├── volcanic activity_json/All.json
├── wildfire/
└── wildfire_json/All.json

Satellite Sensors

The dataset includes multispectral imagery from three satellite sensors:

Sensor	Bands	Description
AMSU-A	15 bands (0-14)	Advanced Microwave Sounding Unit-A
HIRS	20 bands (0-19)	High-resolution Infrared Radiation Sounder
MHS	5 bands (0-4)	Microwave Humidity Sounder

Weather Station Data

Selected samples include supplementary weather station measurements (temperature, pressure, humidity, wind speed) when available for the event location.

VQA Sample Fields

Field	Type	Description
Question_id	string	Unique identifier (format: Task/Type/timestep/subtype/event_id)
Task	string	Task category (Early Warning / Impact Assessment / Recovery Assessment)
Subtask	string	Subtask with timestep, e.g. "Risk Detection (t1)"
Text	string	Question text
Image	string	Comma-separated relative image paths
Stations	object	Weather station sensor data (optional)
Ground truth	string	Standard answer

QID Format

Task/Type/timestep/subtype/event_id

Examples:

• Early Warning/Risk Detection/t1/ground movement/2014-0049-GRC
• Impact Assessment/Magnitude/tmax/tsunami/2011-0282-JPN
• Recovery Assessment/Recovery Time Prediction/t15/cold-wave1/2011-0105-MEX

Evaluation

The evaluation framework supports multiple scoring rules per subtask type:

• Boolean: Exact match (Risk Detection)
• Numeric: Error-based scoring with tolerance (CPI, Magnitude, population counts)
• Classification: Exact or partial match (Type Classification)
• Text: Jaccard word overlap (Arrival Time, Duration, Recovery Time)

Evaluation code is available at: https://github.com/YYQ898/ObsCrisis-Bench

Citation

If you use this dataset, please cite:

@misc{obscrisis-bench,
  title={ObsCrisis-Bench: A Multimodal Benchmark for Extreme Weather Event Analysis},
  author={ObsCrisis Team},
  year={2025}
}

License

MIT License