Source-linked AI summary
Keep Your Friends Close, and the Right Neighbours Closer: Disaster-Conditioned Kernel-Regularized Graph Attention for Building Damage Classification
Fuad Hasan, Chul Min Yeum
TL;DR
Building-damage classifiers need spatial context, but disaster-dependent clustering makes fixed or unconstrained neighbourhood aggregation risky. The paper uses disaster-type-conditioned kernel-regularized graph attention with residual Moran’s I regularization in controlled xBD experiments, reporting improved transfer and reduced residual spatial autocorrelation under shift.
Problem
Building-damage models often rely on per-building appearance despite spatially structured errors and disaster types having different neighbourhood-scale clustering.
Method
The method combines PPC building patches and GPS-derived graphs with a disaster-type-conditioned multi-scale kernel prior in graph attention and a residual Moran’s I penalty.
Results
Under zero-shot event and dataset shift, the proposed model improves over patch-only and vanilla GAT comparisons while reducing residual spatial autocorrelation.
Takeaways & Limitations
The results support treating spatial context as a conditional cue rather than a universal smoothing prior for building-damage classification.
Takeaways & Limitations
The method assumes candidate building instances and centroids from provided polygons, while GPS-only k-NN graphs ignore scene semantics and may connect weakly coupled buildings.
Abstract
from arXiv · showhide
Disaster damage is spatial: buildings rarely fail in isolation. Yet using spatial context for damage classification remains surprisingly underexplored, and many pipelines still rely primarily on per-building appearance cues even when the dominant uncertainty is spatially structured. Complicating matters, the right neighbourhood is not the same across events. Floods, hurricanes, and wildfires can exhibit very different clustering behaviour, making spatial reasoning valuable but easy to misuse - naive context aggregation can improve visual coherence while oversmoothing boundaries or propagating structured errors. We study this tension on xBD (the dataset used in the xView2 challenge) in a controlled post-localization, classification-only setup: each building is represented by a pre/post combined (PPC) patch cropped from the provided polygons, and spatial context is modelled with GPS-derived building graphs. Our approach keeps local evidence "close" by preserving strong spatial relationships in disaster damage patterns, while bringing only the right neighbours "closer" through a disaster-type-conditioned graph model that injects a learnable multi-scale spatial kernel prior into attention, allowing the effective neighbourhood scale to adapt across disaster types rather than being learned as a single global smoothing rule. To discourage coherence-by-smoothing, we add a residual de-correlation loss that penalizes positive Moran's~I in prediction residuals. We evaluate the method under event and dataset shift with a leave-one-event-out (LOEO) protocol on xBD and cross-dataset transfer from xBD to Ida-BD. The model improves macro-F1 and substantially reduces residual spatial autocorrelation under zero-shot event shift, indicating better use of spatial context rather than naive smoothing and enabling more reliable transfer to unseen events within known disaster types.
1 Introduction
Building damage is spatially structured but disaster-dependent, so fixed or unconstrained neighbourhood aggregation can oversmooth boundaries and propagate errors. The paper proposes conditioning graph context on disaster type and evaluates transfer under event and dataset shift.
- Motivation: Independent patch classifiers can overfit event-specific appearance, while spatial models may blur boundaries or propagate neighbourhood errors.These issues limit cross-event generalization and produce spatially clustered residuals.
- Motivation: Hurricanes and wildfires exhibit stronger average spatial autocorrelation than floods, motivating disaster-type-adaptive neighbourhood scales.Longer-range neighbours may be informative when correlation decays slowly, but should not be pulled closer otherwise.
- Study design: The controlled study isolates relational context after localization using PPC building patches and GPS-derived graphs from provided building polygons.The setup tests classification under event shift without confounding detection or segmentation.
- Method: The method combines a disaster-type-conditioned multi-scale kernel prior for graph attention with a residual de-correlation loss based on Moran’s I.The kernel adapts effective neighbourhood scale, while the residual term penalizes clustered prediction errors.
- Method: The kernel-regularized attention layer injects a disaster-type-conditioned spatial prior into GAT-style attention logits.This makes spatial influence distance- and disaster-type-aware rather than globally fixed.
- Evaluation: The evaluation uses xBD leave-one-event-out shift and xBD→Ida-BD cross-dataset shift to stress-test zero-shot context modelling.The protocols assess transfer to held-out events and a separate dataset.
2 Related Work
Prior work establishes building-damage benchmarks, graph-based spatial reasoning, and spatial-autocorrelation diagnostics. This paper combines these strands by using a learnable spatial prior during attention and residual autocorrelation during training.
- Building damage assessment: xBD and xView2 provide large-scale bitemporal imagery, polygon annotations, and four building-level damage categories for damage assessment.Related methods include localization-plus-classification pipelines and bitemporal fusion networks.
- Graph-based context modelling: GraphSAGE, GCN, and GAT-family models provide increasingly expressive message passing, but unconstrained attention can become unstable or oversmooth under event shifts.Prior building-damage work reported benefits alongside oversmoothing-like failures in boundary-heavy scenes.
- Geostatistics and spatial diagnostics: Moran’s I and Geary’s C are standard spatial-autocorrelation diagnostics that can characterize residual spatial structure.The paper extends this perspective by learning a spatial kernel in attention and penalizing residual autocorrelation during training.
3 Method
The method classifies building damage from PPC patches while adding GPS-derived graph context through disaster-conditioned, kernel-regularized attention and residual spatial de-correlation.
- Task setup and training: The pipeline encodes PPC building patches, constructs GPS-derived kNN graphs, performs graph reasoning, and jointly optimizes classification, ordinal, and residual Moran objectives.The graph uses fixed building instances and geometry while updating node representations through message passing.
- Task setup and training: Each building is represented by a pre/post crop from its annotated polygon, with polygon area and perimeter concatenated to the pretrained patch embedding.The PPC tensor is formed by cropping and resizing paired pre/post RGB regions.
- Graph construction and attention: The graph topology and geometric features remain fixed, while attention learns edge weights over the kNN neighbourhood during message passing.Edges store centroid distance and relative direction; node representations and induced attention weights are updated by the network.
- Disaster-conditioned kernel prior: A disaster-type embedding predicts mixture weights and length scales for a three-scale DCMS kernel covering short-, medium-, and long-range context.The mixture weights are shared across an event’s edges, while the resulting prior remains edge-specific through distance.
- Kernel-regularized graph attention: The DCMS prior enters attention logits as a regularizer, retaining data-adaptive attention while anchoring neighbour weighting to an interpretable disaster-type-aware spatial prior.The same fixed neighbourhood induces dynamic, layer-dependent weighted graphs.
- Losses: Training combines class-balanced cross-entropy, an EMD-style ordinal loss, and a residual Moran penalty that discourages positive residual spatial autocorrelation.Residuals are computed from severity labels and predictions using a fixed row-normalized spatial weight matrix.
4 Experiments
The experiments compare fixed-instance classification and graph variants across official holdout, LOEO event shift, sensitivity checks, transfer to Ida-BD, and spatial diagnostics. The full model improves classification while reducing residual spatial autocorrelation under the stricter LOEO setting.
- Evaluation protocols: The study uses three evaluation protocols: official xView2 holdout, zero-shot LOEO cross-event transfer, and zero-shot xBD-to-Ida-BD transfer.LOEO holds out one event at a time and provides its disaster type as metadata; the xView2 row is an external reference rather than a direct comparison.
- xView2 holdout: 0.87288 Macro-F1 is achieved by the full model under the controlled xView2 holdout, versus 0.84102 for Vanilla GAT.The comparison fixes building instances and evaluates the classification/context module using the same PPC patch protocol.
- LOEO stress test: 0.503 LOEO macro-F1 and 0.079 residual Moran’s I are achieved by the full model, compared with 0.453 and 0.256 for vanilla GAT.The patch-only baseline reaches macro-F1 0.433, while the full model provides the strongest joint accuracy and spatial-consistency result.
- Ablations: The kernel prior provides the strongest standalone gain, while disaster conditioning improves performance consistently and the residual term mainly reduces spatially structured error.Pairwise ablations show kernel + residual outperform either component alone, with disaster conditioning producing the best final result.
- Confusion analysis: The proposed model increases diagonal concentration across LOEO confusion matrices while retaining difficult Minor↔Major and Major↔Destroyed confusions.Vanilla GAT shifts some errors toward adjacent-class leakage consistent with oversmoothing-like behaviour in mixed neighbourhoods.
- Spatial and transfer diagnostics: Learned kernel mixtures vary across disaster rollups, and the proposed model generally improves rollup-level Macro-F1 while lowering residual Moran’s I.Attention-distance diagnostics also show stronger same-label versus different-label distance structure than vanilla GAT; xBD-to-Ida-BD transfer follows the same improvement trend.
5 Discussion
The discussion argues that disaster-type-conditioned spatial context improves damage classification under event and dataset shift, while reducing risks from unconstrained smoothing. It also identifies important scope boundaries around localization, graph semantics, metadata, and sensing modalities.
- 5 Discussion: LOEO is the diagnostic setting because changing event conditions produce spatially clustered residuals in patch-only predictions.The fixed-instance protocol isolates whether spatial context improves classification rather than localization.
- 5 Discussion: Vanilla GAT improves over patch-only classification, but unconstrained attention can propagate mistakes across mixed neighbourhoods.The proposed DCMS kernels condition context radius on disaster type, while residual Moran regularization discourages coherent-but-wrong errors.
- 5 Discussion: Spatial context is a conditional cue: clustered labels can make neighbours informative, whereas forcing neighbourhood agreement under weak clustering can introduce bias.This frames disaster-specific spatial scale as preferable to a universal smoothing prior.
- 5 Discussion: Cross-dataset xBD→Ida-BD transfer shows the proposed model improving over patch-only and vanilla GAT, with gains concentrated in Minor and Major classes.The comparison uses fixed building instances and a patch protocol despite differences in published Ida-BD baselines.
- 5 Discussion: The method assumes candidate instances and centroids, uses GPS-only k-NN graphs, requires known disaster type, and relies on optical PPC patches.Detector-generated footprints, topology- or uncertainty-aware edges, inferred metadata, and additional modalities remain future directions.
6 Conclusion
The paper presents a disaster-type-conditioned, kernel-regularized graph attention model for building damage classification in a controlled post-localization xBD setting. Across event and dataset shift, it improves macro-F1 and reduces residual spatial autocorrelation.
- 6 Conclusion: The method combines PPC building patches, GPS-derived graph structure, disaster-type-adaptive multi-scale kernels, and residual Moran regularization.This makes neighbourhood influence distance- and type-aware while discouraging coherent-but-wrong residual errors.
- 6 Conclusion: Across xBD LOEO and xBD→Ida-BD evaluations, the model improves macro-F1 and reduces residual Moran’s I.These evaluations cover zero-shot event shift and dataset shift in the controlled classification setting.