---
pretty_name: The CRASAR sUAS Disaster Response Overhead Inspection Dataset
size_categories:
- 100K<n<1M
---
# The CRASAR sUAS \[D\]isaster \[R\]esponse \[O\]verhead \[I\]nspection \[D\]ata\[s\]et

This repository contains the CRASAR-U-DROIDs dataset. This is a dataset of orthomosaic images with accompanying labels for building damage assessment. The data contained here is described in [this document](https://arxiv.org/abs/2407.17673).

This dataset contains 52 orthomosaics containing 21716 building polygons collected from 10 different disasters, totaling 67 gigapixels of imagery. 
Building polygons were sourced from Microsoft's US Building Footprints Dataset \[[1](https://github.com/microsoft/USBuildingFootprints)\], and in some cases, building polygons were added manually by the authors. 
Each building polygon has been annotated using the Joint Damage Scale \[[2](https://arxiv.org/abs/1911.09296)\] and translationally aligned for model training.
The dataset has been split into test and train at the disaster level, with 6 disasters in the train set and 4 disasters in the test set.
A summary of the dataset, grouped by disaster and ordered by area, is included below for reference.

| Disaster                  | Area (km^2)     | Gigapixels      | Building Labels   | Orthomosaics  | Test or Train |
| ------------------------- | --------------- | --------------- | ----------------- | ------------- | ------------- |
| Hurricane Ian             | 32.66517523     | 33.19155902     | 100351            | 200           | Train         |
| Mayfield Tornado          | 8.422144185     | 9.698707535     | 2028              | 3             | Test          |
| Kilauea Eruption          | 5.751864646     | 1.121020488     | 382               | 3             | Train         |
| Hurricane Idalia          | 5.686794335     | 1.095231308     | 4636              | 12            | Test          |
| Hurricane Ida             | 5.139696352     | 6.976915134     | 2068              | 9             | Train         |
| Hurricane Michael         | 3.617024461     | 9.567229047     | 6859              | 12            | Test          |
| Hurricane Harvey          | 2.596253635     | 5.128525423     | 5546              | 17            | Train         |
| Hurricane Laura           | 2.341867225     | 1.456463        | 500               | 3             | Train         |
| Mussett Bayou Fire        | 1.714575473     | 2.164129413     | 128               | 5             | Test          |
| Champlain Towers Collapse | 0.041536185     | 0.246084846     | 4                 | 1             | Train         |
| **Total**                 | **67.97693173** | **70.64586393** | **122502**        | **265**       | **N/A**       |

## Dataset Structure
At the top level, the dataset contains a statistics.csv file, with summary statistics of the dataset, and two folders, train and test. 
Each folder has imagery (which contains all of the geo.tif files) and annotations.
The annotations folder then contains one folder for each source of imagery (and therefore labels): sUAS, SATELLITE, and CREWED. These folders contain the imagery-derived labels from the imagery associated with each of the imagery sources.
These folders contain two groups of data: alignment_adjustments, and building_damage_assessment.
These two groups of data contain JSON data that represent the annotations for both building damage assessment and the translational alignments necessary to align the building polygons with the imagery.
These two data sources are discussed below.

### Building Damage Assessment
A sample of the building damage assessment JSON file is as follows... 
```json
{
  "8c6ab8368b726ace3807f7e64cceceb8": [
    {
      "EPSG:4326": [
        {
          "lat": 30.096003,
          "lon": -93.727638
        },
        {
          "lat": 30.096012,
          "lon": -93.727236
        },
        {
          "lat": 30.09609,
          "lon": -93.727238
        },
        {
          "lat": 30.096082,
          "lon": -93.72764
        },
        {
          "lat": 30.096003,
          "lon": -93.727638
        }
      ],
      "id": "03d519a729349cb55d581b145a780dd3",
      "label": "no damage",
      "pixels": [
        {
          "x": 1116,
          "y": 1128
        },
        {
          "x": 1414,
          "y": 1121
        },
        {
          "x": 1412,
          "y": 1063
        },
        {
          "x": 1114,
          "y": 1069
        },
        {
          "x": 1116,
          "y": 1128
        }
      ],
      "source": "Microsoft",
      "boundary": "0827-B-02.geo.tif"
    },
    ...
  ],
  ...
}
```

Each JSON file is a dictionary that contains entries where the key is a building id, and each value is a list of all of the instances (views) where that building appears in the dataset's imagery. Each view contains the following information...
- The "source" field describes the provenance of the building polygon. The possible options are "Microsoft," indicating the building polygon was sourced from the Microsoft Building Footprints dataset, and "custom," indicating the polygons were manually added by the authors.
- The "id" field is a string that uniquely identifies each building. Building polygons that appear across multiple orthomosaics will share the same id. Each id will only appear once in each orthomosaic.
- The "label" field corresponds to the values of the Joint Damage Scale. The possible options are "no damage", "minor damage", "major damage", "destroyed", and "un-classified".
- The "pixels" field corresponds to the coordinates of the building polygon in the pixel coordinate space of the orthomosaic.
- The "EPSG:4326" field corresponds to the coordinates of the building polygon in the EPSG:4326 coordinate space.

### Alignment Adjustments
A sample of the alignment adjustment JSON file is as follows...
```json
[[[4739.728, 4061.728], [4542.137, 3962.933]], ... ]
```

Each JSON file is a list of lines with a length of two, each defined by a 2d coordinate corresponding to an x,y pixel coordinate in the orthomosaic.
The first list represents a list of all the alignment adjustments for the given orthomosaic.
The second list represents a set of two points, forming a line, that describes the translational adjustment needed to bring the nearby building polygons into alignment with the imagery.

Each translational adjustment starts with the position in the unadjusted coordinate space that needs to be moved to the following position in order to align the building polygons.
These translational adjustments are applied to the building polygons by applying the nearest adjustment to each building polygon. 
Functionally, this forms a vector field that describes the adjustments for an entire orthomosaic.
This process is described in detail in [3](https://arxiv.org/abs/2405.06593).

## Publications & Documentation

The following papers exist that describe the dataset and its intended uses...

1) [CRASAR-U-DROIDs: A Large Scale Benchmark Dataset for Building Alignment and Damage Assessment in Georectified sUAS Imagery](https://arxiv.org/abs/2407.17673). This paper presents the Center for Robot Assisted Search And Rescue - Uncrewed Aerial Systems - Disaster Response Overhead Inspection Dataset (CRASAR-U-DROIDs) for building damage assessment and spatial alignment collected from small uncrewed aerial systems (sUAS) geospatial imagery. To replicate the results from this paper, please see commit ae3e394cf0377e6e2ccd8fcef64dbdaffd766434.
2) [\[FAccT'25\] Now you see it, Now you don’t: Damage Label Agreement in Drone & Satellite Post-Disaster Imagery](). This paper audits damage labels derived from coincident satellite and drone aerial imagery for 15,814 buildings across Hurricanes Ian, Michael, and Harvey, finding 29.02\% label disagreement and significantly different distributions between the two sources, which presents risks and potential harms during the deployment of machine learning damage assessment systems. To replicate the results of this paper, please use the source code located [here](https://github.com/TManzini/NowYouSeeItNowYouDont/), and the data found at commit 58f0d5ea2544dec8c126ac066e236943f26d0b7e.
3) [\[RO-MAN'25\] Non-Uniform Spatial Alignment Errors in sUAS Imagery From Wide-Area Disasters](https://arxiv.org/abs/2405.06593). This work presents the first quantitative study of alignment errors between small uncrewed aerial systems (sUAS) geospatial imagery and a priori building polygons and finds that alignment errors are non-uniform and irregular. To replicate the results from this paper, please see commit ae3e394cf0377e6e2ccd8fcef64dbdaffd766434.
4) [\[RO-MAN'25\] Challenges and Research Directions from the Operational Use of a Machine Learning Damage Assessment System via Small Uncrewed Aerial Systems at Hurricanes Debby and Helene](https://arxiv.org/pdf/2506.15890). This work presents the research directions and challenges that were encountered from the operational deployment of ML models trained on the CRSAR-U-DROIDs dataset. To find the data used to train the models used in this work, please see commit ae3e394cf0377e6e2ccd8fcef64dbdaffd766434.