CRASAR-U-DRIODs: Post-Disaster Imagery Benchmark

• CRASAR-U-DRIODs is a high-resolution sUAS benchmark offering rigorously aligned annotations for over 21,700 building footprints and 657 km of road segments across diverse disasters.
• It leverages multiple drone platforms and advanced orthomosaic processing to achieve precise georectification and tile-based annotation for robust machine learning validation.
• The dataset enhances damage assessment and rapid response strategies by providing actionable insights for post-disaster infrastructure repair and emergency decision-making.

The Center for Robot Assisted Search And Rescue Uncrewed Aerial Systems Disaster Response Overhead Inspection Dataset (CRASAR-U-DRIODs) comprises a large-scale, high-resolution benchmark for building alignment and damage assessment in post-disaster environments. Acquired entirely from small uncrewed aerial systems (sUAS) across multiple disaster types, CRASAR-U-DRIODs is distinguished by its rigorous spatial alignment, comprehensive human annotation via a layered review process, and interoperability with satellite-oriented benchmarks. It serves as the de facto reference for training and evaluating machine learning systems on high-resolution sUAS imagery, enabling robust assessment of both building and road infrastructure damage over real-world events (Manzini et al., 2024, Manzini et al., 13 Dec 2025).

1. Dataset Composition and Geospatial Coverage

CRASAR-U-DRIODs comprises 52 georectified orthomosaics collected during ten federally declared U.S. disasters, spanning a cumulative area of 67.98 km² (26.245 mi²). These events encompass diverse disaster scenarios, including hurricanes (Harvey, Laura, Michael, Ian, Ida, Idalia), wildfire (Musset Bayou Fire), tornado outbreak (Mayfield, KY), volcanic eruption (Kīlauea), and urban structural collapse (Champlain Towers) (Manzini et al., 2024). The dataset provides annotations over 21,716 building footprints sourced from Microsoft Building Footprints, subsequently manually spatially adjusted, and 7,880 unique adjustment vectors for corrective translation. Associated datasets extend to spatially aligned road labels, with 657.25 km of post-disaster roadways labeled under a 10-class scheme (Manzini et al., 13 Dec 2025).

Dimension	Buildings Subset	Roads Subset
Number of orthomosaics	52	52
Annotated objects	21,716 building polygons, 7,880 adjustments	657.25 km of road segments
Hazard types included	Wind, flood, fire, volcanic, collapse	Wind, flood, fire, tornado
Area covered	67.98 km²	≥67.98 km² (coincident)

2. Acquisition, Orthomosaic Processing, and Tiling

Imagery acquisition leveraged a heterogeneous fleet, including DJI (Mavic Pro, Mavic 2, M300/M30T, Phantom 4, M600), SenseFly eBee X, Parrot Anafi, and WingtraOne Gen II platforms. Flights were conducted predominantly at low-altitude, resulting in mean ground sample distance (GSD) values of 3.74 cm/px (range: 1.77–12.7 cm/px); these consumer-grade RGB images were subsequently processed into orthomosaics via Pix4D React (50 of 52 mosaics) or Agisoft Metashape (2 of 52) (Manzini et al., 2024). Final outputs utilized GeoTIFF with embedded georeferencing; each orthomosaic was partitioned for annotation into 2048×2048 px tiles (full detail) and 8500×8500 px tiles (bulk "no-damage" processing), with 5% overlap to prevent annotation gaps.

For road assessment, orthomosaics were similarly tiled—45 at 2048 px and 7 at 8500 px—and OpenStreetMap linework was intersected and buffered (7.2 m-wide corridors) for damage annotation (Manzini et al., 13 Dec 2025). Polygons were derived by algorithmic expansion along each road segment to ensure consistent label rasterization across resolutions.

3. Spatial Alignment: Methodology and Quality

Primary sources of misalignment derive from GPS noise (sUAS telemetry), disparities between satellite-derived versus sUAS GSDs, and residual orthorectification artifacts. To address these, CRASAR-U-DRIODs employs annotation-driven alignment correction: for each misaligned building, a human-provided translational vector corrects spatial positioning, yielding 7,880 adjustment annotations (≈36% of buildings). These vectors are interpolated to produce dense, per-orthomosaic displacement fields; all building polygons are then shifted accordingly (Manzini et al., 2024). For roads, 9,184 vertex adjustments were similarly applied, guaranteeing that 100% of annotated vertices align within the prescribed 7.2 m corridor (Manzini et al., 13 Dec 2025).

Assessment of residual spatial error is supported via standard root-mean-square error (RMSE) on sampled control points: $$\mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| x^{\mathrm{pred}}_i - x^{\mathrm{true}}_i \|^2 }$$ A plausible implication is that alignment correction substantially mitigates label/feature mis-registration, elevating downstream model performance and facilitating accurate cross-asset and multi-modal benchmarking.

4. Annotation Protocol and Damage Classification Schema

Human annotation, executed via the Labelbox platform, comprised a pool of 130 annotators. For buildings, 55 annotators labeled 18,780 tiles at 2048 px, while three dataset authors provided bulk "no-damage" labeling for coarse tiles. The annotation protocol followed a two-stage QC process: initial expert reviewer correction affecting 10.4% of tiles, followed by committee audit yielding corrections to 5.9% of building labels and addition of 0.5% new building footprints (Manzini et al., 2024). For roads, 130 annotators contributed across 17,600 tiles; spot QC (9.7% tiles, 12% road length) ensured consistent standard.

Building damage is classified via the Joint Damage Scale (JDS), interoperable with xBD, comprising five categories:

1. No damage
2. Minor damage
3. Major damage
4. Destroyed
5. Un-classified (building missing/destruction uncertain)

Road damage adopts a ten-class taxonomy, developed with FEMA and TxDOT input, including fine-grained distinctions across obstructions, flood, particulate coverage, and structural obliteration (Manzini et al., 13 Dec 2025).

Inter-annotator agreement statistics are not explicitly reported, but can be formally evaluated using Cohen’s κ-statistic: $\kappa = \frac{p_o - p_e}{1 - p_e}$ where $p_o$ denotes observed agreement and $p_e$ chance agreement.

5. Data Structure, Distribution, and Licensing

The dataset is organized in a comprehensive directory structure, with top-level /train/ and /test/ partitions. Each orthomosaic subfolder contains:

• {YYYYMMDD}-{Location}.geo.tif: GeoTIFF image
• {...}.buildings.shp or GeoJSON: Aligned building footprints + JDS labels
• {...}.adjustments.csv: Per-vertex translation annotations
• {...}.metadata.json: GSD, CRS (EPSG code), collection date, event name, platform type, annotator counts

Corresponding road data includes linework and label rasters, adjustment vectors, and event-level metadata. The public release is hosted at https://huggingface.co/datasets/CRASAR/CRASAR-U-DROIDs; raw source imagery is available on request for provenance. Licensing is CC-BY-4.0 (Manzini et al., 2024).

6. Benchmarking, Applications, and Limitations

CRASAR-U-DRIODs enables a spectrum of machine learning research: building damage assessment, spatial alignment benchmarking, robust multi-scale model training, and cross-platform transfer between sUAS and satellite datasets. Baseline models on the affiliated road damage subset encompass 18 configurations (nine architectures × two tasks: three-class vs. ten-class), including UNet, Attention UNet, ResNet101+PSPNet, DeepLabV3+, ViT-L Segmenter/UperNet (pretrained and from scratch) (Manzini et al., 13 Dec 2025).

Key performance metrics include mean Intersection-over-Union (mIoU) and per-class accuracy, calculated on physical extent ($\mathrm{IoU}_{\mathrm{km}}$). The top-performing model for the road “Simple” task is Attention UNet with Macro IoU$_{\mathrm{km}}$ = 0.331, while the “Full” (ten-class) task is led by ResNet101+DeepLabV3+ with Macro IoU$_{\mathrm{km}}$ = 0.091. Spatial alignment improves average Macro IoU by approximately 5.6%. The building damage dataset is released for benchmarking with U-Net, Mask R-CNN, and transformer-based architectures, though no building-damage-specific CV/ML benchmarks are reported in the primary publication (Manzini et al., 2024).

Operational usage has included deployments on Hurricanes Debby and Helene, producing actionable overlay products (KML) for field response in near real time. Feedback indicates that false positives are less problematic than false negatives in emergency contexts, and that full-schema outputs are desired for downstream routing and engineering prioritization (Manzini et al., 13 Dec 2025).

Current limitations include class imbalance, underperformance on rare or subtle subclasses (partial road conditions), and difficulties generalizing across disaster types. Open challenges comprise improved learning of infrequent damage conditions, end-to-end automation of spatial alignment, domain shift adaptation, and ethical considerations arising from potential misclassifications.

7. Significance and Outlook

CRASAR-U-DRIODs sets a large-scale precedent for post-disaster infrastructure assessment from sUAS imagery, bridging the operational gap between raw geospatial data and practitioner-applicable products. Its precise multi-stage spatial alignment, high-fidelity labeling, and class taxonomies coherent with satellite datasets provide a robust foundation for reproducible benchmarking and transfer learning. This suggests near-term utility not only for ML/CV model evaluation but also for advancing spatial alignment algorithms and informing rapid-response robotics and emergency decision-support systems. Future directions may include richer multi-modal annotations (e.g., temporal change detection), integration with automated georeferencing pipelines, and formal evaluation protocols for end-to-end system robustness across unseen events (Manzini et al., 2024, Manzini et al., 13 Dec 2025).