The Potential of Copernicus Satellites for Disaster Response: Retrieving Building Damage from Sentinel-1 and Sentinel-2

Abstract: Natural disasters demand rapid damage assessment to guide humanitarian response. Here, we investigate whether medium-resolution Earth observation images from the Copernicus program can support building damage assessment, complementing very-high resolution imagery with often limited availability. We introduce xBD-S12, a dataset of 10,315 pre- and post-disaster image pairs from both Sentinel-1 and Sentinel-2, spatially and temporally aligned with the established xBD benchmark. In a series of experiments, we demonstrate that building damage can be detected and mapped rather well in many disaster scenarios, despite the moderate 10$\,$m ground sampling distance. We also find that, for damage mapping at that resolution, architectural sophistication does not seem to bring much advantage: more complex model architectures tend to struggle with generalization to unseen disasters, and geospatial foundation models bring little practical benefit. Our results suggest that Copernicus images are a viable data source for rapid, wide-area damage assessment and could play an important role alongside VHR imagery. We release the xBD-S12 dataset, code, and trained models to support further research.

• The paper introduces xBD-S12, a novel dataset of 10,315 paired images that aligns Sentinel-1/2 imagery with VHR data for building damage detection.
• The study employs a lightweight U-Net with a ResNet-34 backbone and dual fusion strategies to enhance multi-modal, bi-temporal segmentation performance.
• The analysis reveals that Sentinel imagery is effective for extensive, spectrally distinct disasters but is limited by 10 m resolution in sparse urban areas.

The Potential of Copernicus Satellites for Disaster Response: Retrieving Building Damage from Sentinel-1 and Sentinel-2

Introduction and Motivation

The paper presents a rigorous investigation into the utility of publicly available medium-resolution Copernicus satellite imagery (Sentinel-1 SAR, Sentinel-2 multispectral) for automated building damage assessment following natural disasters. The underlying motivation derives from the limited accessibility and coverage of VHR ($\leq$2\,m GSD) commercial satellite data during critical first-response periods, especially for large-scale disaster events. The work introduces xBD-S12, a substantial dataset of pre- and post-disaster image pairs that are temporally and spatially matched to the canonical VHR xBD benchmark. This enables direct supervised learning and evaluation of damage classification at Sentinel-level resolution.

Figure 1: Copernicus satellites are surprisingly effective at mapping building damage, despite the moderate GSD of 10\,m.

Dataset Construction and Properties

A central contribution is the creation of xBD-S12, comprising 10,315 image pairs across 16 disasters and multiple event types, curated by aligning xBD VHR images with corresponding Sentinel-1 and Sentinel-2 tiles. The authors implement systematic georeferencing corrections to resolve calibration shifts in the xBD images, employing building footprint polygons as ground control points for robust affine rectification. The dataset includes rigorous spatial resampling (fixed $128\times128$ px, $\sim$4\,m GSD), cloud masking using CloudScore+, and morphological buffering to accommodate misregistration and labeling uncertainties.

Figure 2: xBD-S12 comprises 10,315 image pairs across 16 disaster events; colors indicate the event-based test split.

Figure 3: Geographical distribution of xBD-S12, grouped by disaster type.

Sentinel-2 inputs are selected through a dual optimization over cloud coverage and disaster temporal proximity, while Sentinel-1 acquisitions are matched for orbit direction and date. All dataset patches are processed to harmonize input dimensions, with label simplification to three classes: background, intact, and damaged—reflecting the inherent ambiguity and spatial resolution constraints (where individual buildings occupy few pixels).

Figure 4: Example patches showing the Sentinel-2 TCI, Sentinel-1 amplitude, and reference VHR images plus damage labels.

Model Architecture and Training Strategy

Building damage is formulated as a pixel-level multiclass segmentation task across bi-temporal, bi-modal Sentinel imagery. The authors implement a lightweight U-Net with a ResNet-34 backbone (ImageNet pretraining), with ablation studies demonstrating that encoder depth and removal of batch normalization layers optimize practical performance.

Fusion strategies are evaluated—early fusion (channelwise concatenation) and late fusion (Siamese encoding of pre-/post- images)—with both single-joint and two-step (sequential localization and damage classification) optimization pipelines. The two-step pipeline incorporates morphological dilation (3-pixel buffer) to prioritize recall over boundary precision, a critical consideration at moderate satellite resolution.

Training employs progressive class balancing, geometric augmentations, and biased sampling to address extreme class imbalance (background and intact vastly outnumber damaged), with standard cross-entropy loss outperforming weighted alternatives given the simplified labels.

Empirical Results and Analysis

Comprehensive experimentation is presented on both the canonical xView2 split (random across events) and an event-based split (geographic segregation), with metrics reported for F1 localization, F1 damage, and a weighted composite “competition score.” The key empirical finding is that simple architectures (Siamese U-Net, 2-step) exhibit superior generalization to unseen disasters under event-based splitting, while more sophisticated methods (e.g., ChangeMamba) show better in-domain performance but attenuated transferability.

Figure 5: Effect of buffer size on building localization and corresponding F1 scores—larger buffers increase recall at the cost of detail.

Building damage detection is robust for events with contiguous, spectrally distinct footprints (wildfires, floods), with competition F1 scores exceeding 0.8. Hurricane-induced and earthquake-induced damages, which are spatially sparse and challenging at 10\,m resolution, show substantially lower scores, indicating a clear limitation of Sentinel imagery for fine-grained urban damage.

Interestingly, late-fusion models and multi-temporal inputs improve localization and generalization in unseen geographic regions, suggesting temporal redundancy and atmospheric invariance are beneficial.

Comparative Evaluation with Geospatial Foundation Models

Two recent foundation models—Prithvi-EO-2.0 and DOFA—are benchmarked as frozen feature extractors with UperNet decoders. They provide modest improvements in specific scenarios but exhibit pronounced overfitting and poor generalization on the event-based split, with absolute competition scores trailing the tuned U-Net variants. Even in low-data regimes (event-to-event transfer), foundation models do not consistently outperform the baseline, except possibly for broad environmental changes (e.g., fire burn scars). The authors explicitly note that finetuning the encoder may marginally improve results, but at substantially increased computational cost.

Qualitative Generalization

The best-performing model is applied in zero-shot mode to a recent out-of-distribution (2025 Palisades wildfire) event, demonstrating qualitative alignment of predicted damage maps against external government damage assessments despite the lack of VHR ground truth. The output provides a rapid, wide-area overview suitable for field deployment.

Figure 6: Complete test set for the Guatemala volcano disaster and model predictions illustrating the limits of Sentinel resolution.

Figure 7: Extreme class imbalance in the Mexico earthquake test set; damaged pixels comprise only 0.4%.

Limitations and Practical Deployment

Sentinel-based building damage mapping delivers high precision and recall for disasters with spatially extensive, spectrally distinct signals but is structurally limited for sparse, urban damage scenarios. The primary bottleneck is geolocation accuracy for building footprints and the moderate 10\,m GSD, at which buildings are poorly delineated or indistinguishable from background. The methodology is not an operational substitute for VHR imagery, but is a valuable complement where rapid access or coverage is prioritized.

This work emphasizes simplicity (standard GRD Sentinel products, minimal temporal stack) for practical accessibility and reproducibility, with released code and trained weights facilitating further work. The inclusion of buffer masking and label simplification is shown to be critical for optimizing model recall—a necessary tradeoff in disaster response assessment.

Implications and Future Directions

The explicit empirical finding that architectural sophistication (transformers, Mamba, object-based methods) does not bring consistent advantage at 10\,m GSD, and that GeoFMs yield limited practical benefit for this task, suggests that future work should focus on overcoming spatial resolution and geolocalization limitations rather than network architecture complexity. The authors advocate for operational fusion with national/global building inventories (e.g., Overture Maps) and suggest that Sentinel time series (InSAR, coherence maps) may yield higher sensitivity for subtle damage patterns.

Further, the continued improvement of foundation models (GeoFMs/VLMs) may eventually enable better transfer and task-specific adaptation as pretraining strategies and domains evolve.

Figure 8: Additional example patches from xBD-S12 visualizing Sentinel inputs and VHR reference images plus damage labels.

Conclusion

This study provides robust evidence that Copernicus Sentinel satellite data is a viable and surprisingly effective data source for rapid building damage mapping in many disaster scenarios, despite moderate spatial resolution. The xBD-S12 dataset, architectural findings, and validation experiments set a foundation for scalable models and operational workflows where VHR imagery is unavailable or limited. The results delineate the boundaries of effectiveness, highlight the value of simplicity, and emphasize avenues for future research in temporal modeling and geospatial foundation model finetuning for disaster response.