crumeike/tornadonet-datasets

TornadoNet Dataset: Street-View Building Damage Assessment

Dataset Description

TornadoNet is a comprehensive benchmark dataset for automated post-disaster building damage assessment using street-level imagery. The dataset contains high-resolution geotagged images collected following the December 10-11, 2021 Midwest U.S. tornado outbreak, providing realistic conditions for evaluating modern object detection architectures on multi-level damage classification tasks.

Dataset Summary

• Total Images: 3,333 high-resolution street-view images
• Annotated Instances: 8,890 building instances
• Damage Classes: 5 levels (DS0-DS4) based on IN-CORE classification
• Collection Method: Vehicle-mounted 360° cameras
• Resolution: 4K equirectangular images
• Geographic Coverage: 4 of the 9 U.S. states affected by 2021 Midwest tornado outbreak
• Annotation Standard: IN-CORE (Interdependent Networked Community Resilience Modeling Environment)

Supported Tasks

• Object Detection: Detect and localize damaged buildings in street-view imagery
• Multi-class Classification: Classify building damage severity into 5 ordinal levels
• Damage Assessment: Support automated post-disaster reconnaissance and response

---

Quick Start

Loading the Dataset

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("crumeike/tornadonet-datasets")

# Access specific split
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test"]

Using with YOLO

# Dataset is already in YOLO format
# Annotation format: <class> <x_center> <y_center> <width> <height>

# Example YOLO configuration (data.yaml)
"""
path: /path/to/tornadonet
train: images/train
val: images/val
test: images/test

nc: 5  # number of classes
names: ['DS0_Undamaged', 'DS1_Slight', 'DS2_Moderate', 'DS3_Extensive', 'DS4_Complete']
"""

Evaluation Metrics

For ordinal classification tasks, consider using:

• Standard Metrics: mAP@0.5, F1-score, Precision, Recall
• Ordinal Metrics:
  • Ordinal Top-k Accuracy
  • Mean Absolute Ordinal Error (MAOE)
  • Confusion matrices emphasizing near-miss errors

Baseline Models

Benchmark results on TornadoNet:

Model	mAP@0.5	F1 Score	Ordinal Top-1 Acc	MAOE
YOLOv8n	40.98%	45.11%	84.01%	0.78
YOLOv8l	42.09%	46.41%	84.19%	0.78
YOLO11x	46.05%	49.40%	85.20%	0.76
RT-DETR-L	39.87%	44.77%	88.13%	0.65

Ordinal Supervision Impact

Model	Configuration	mAP@0.5	Δ vs Baseline	Ordinal Top-1	MAOE
RT-DETR-L	ψ=0.5, K=1	44.70%	+4.8 pp	91.15%	0.56

See full paper for detailed experimental results and analysis.

---

Dataset Structure

Data Instances

Each instance consists of:

• Image: High-resolution street-view photograph
• Bounding boxes: YOLO format annotations (class x_center y_center width height)
• Damage class: Integer label (0-4) corresponding to damage severity
• Metadata: Geolocation data (when available)

Data Fields

The annotations follow YOLO format:

<class_id> <x_center> <y_center> <width> <height>

Where:

• class_id: Damage state (0=DS1, 1=DS2, 2=DS3, 3=DS4, 4=DS0,)
• x_center, y_center: Normalized bounding box center coordinates (0-1)
• width, height: Normalized bounding box dimensions (0-1)

Damage State Definitions (IN-CORE Framework):

Class ID	Label	Description	Typical Indicators
0	DS1 - Slight	Minor cosmetic damage	2-15% roof covering damaged, 1 window/door failure
1	DS2 - Moderate	Noticeable damage, repairable	15-50% roof damage, 2-3 windows/doors failed
2	DS3 - Extensive	Severe damage, major repairs needed	>50% roof damage, >3 windows/doors failed, 1-3 roof sheathing sections failed
3	DS4 - Complete	Structural collapse or near-total destruction	>35% roof sheathing failed, roof-to-wall connection failure
4	DS0 - Undamaged	No visible structural damage	Intact roof, windows, walls

Data Splits

Split	Images	Instances	Percentage
Train	~2,500	6,184	75%
Validation	~500	1,342	15%
Test	~500	1,364	15%

Class Distribution (across all splits):

• DS0 (Undamaged): ~45%
• DS1 (Slight): ~25%
• DS2 (Moderate): ~15%
• DS3 (Extensive): ~10%
• DS4 (Complete): ~5%

Note: The dataset exhibits natural class imbalance, with fewer instances of severe damage (DS3-DS4).

Dataset Creation

Curation Rationale

Traditional manual post-disaster damage assessments are:

• Labor-intensive and time-consuming
• Subject to cognitive biases and inconsistencies
• Unsafe for personnel in hazardous areas
• Unable to provide real-time, building-level information

TornadoNet was created to:

1. Enable development of automated damage assessment systems
2. Benchmark modern object detection architectures for disaster response
3. Support research in ordinal classification for severity grading
4. Provide standardized evaluation protocols for damage detection models

Source Data

Initial Data Collection

• Event: December 10-11, 2021 Midwest U.S. tornado outbreak
• Collection Timing: ~3 weeks post-event
• Equipment: Vehicle-mounted GoPro cameras (360° panoramic video)
• Coverage: Prioritized heavily impacted areas based on:
  • Post-event aerial imagery
  • Preliminary damage reports
  • Social vulnerability indices
• Processing: Automatic extraction of building-centered frames using geospatial alignment

Who are the source data producers?

Data collection was conducted by the Center of Excellence for Risk-Based Community Resilience Planning (CoE), a NIST-funded center, as part of a longitudinal field study to support empirical validation of the IN-CORE modeling platform.

Annotations

Annotation Process

1. Manual Annotation: Trained researchers drew bounding boxes around individual buildings
2. Damage Classification: Each instance labeled according to IN-CORE five-level damage classification
3. Quality Control: Two-stage validation process:
   • Initial annotation by trained annotators
   • Secondary expert review for consistency verification
4. Tools: LabelImg and custom annotation interfaces
5. Guidelines: Archetype-specific indicators for 19 structural archetypes (T1-T19), with majority being residential wood-frame structures (T1-T5)

Who are the annotators?

Annotations were performed by trained researchers familiar with structural engineering and disaster damage assessment, following standardized IN-CORE guidelines. All annotations underwent expert cross-validation.

Personal and Sensitive Information

The dataset contains street-view imagery of buildings in disaster-affected areas. While efforts were made to focus on structural damage:

• No personally identifiable information (PII) was intentionally collected
• Images may incidentally capture public spaces, vehicles, or street scenes
• Geographic metadata is included for research purposes
• Researchers should use appropriate care when publishing derived visualizations

Additional Information

Dataset Curators

TornadoNet was curated by researchers from:

• Johns Hopkins University
• University of Alabama
• University of South Alabama

In collaboration with the Center of Excellence for Risk-Based Community Resilience Planning (NIST-funded).

Licensing Information

License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)

This dataset is released for research and educational purposes. Commercial use requires separate permission.

Citation Information

If you use this dataset, please cite:

@article{umeike2026tornadonet,
  title={TornadoNet: Real-Time Building Damage Detection with Ordinal Supervision},
  author={Umeike, Robinson and Pham, Cuong and Hausen, Ryan and Dao, Thang and Crawford, Shane and Brown-Giammanco, Tanya and Lemson, Gerard and van de Lindt, John and Johnston, Blythe and Mitschang, Arik and Do, Trung},
  journal={arXiv preprint arXiv:2603.11557},
  year={2026}
}

Contributions

Dataset collection and annotation were supported by:

• Center of Excellence for Risk-Based Community Resilience Planning (NIST Cooperative Agreement 70NANB15H044)
• SciServer computational resources (NSF Award ACI-1261715)

Related Resources

• Code Repository: https://github.com/crumeike/TornadoNet
• IN-CORE Platform: https://incore.ncsa.illinois.edu/
• Project Documentation: https://arxiv.org/abs/2603.11557

Contact

For questions, issues, or collaboration opportunities:

• GitHub Issues: https://github.com/crumeike/TornadoNet/issues
• Dataset Maintainer: crumeike@crimson.ua.edu

Updates and Versions

Version 1.0 (Initial Release)

• 3,333 images with 8,890 annotated instances
• Train/Val/Test splits (75%/15%/15%)
• IN-CORE 5-level damage classification

---

Acknowledgments: We thank the affected communities, first responders, and all those who contributed to disaster response and recovery efforts.